Create OptionalSwitch node + EmptyInputSentinel to deal with inputs that should be seen as 'empty' when passed into nodes

* Fix bypass dtype/device moving

* Force offloading mode for training

* training context var

* offloading implementation in training node

* fix wrong input type

* Support bypass load lora model, correct adapter/offloading handling

comfy/ldm/ace/ace_step15.py

@@ -7,6 +7,67 @@ from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
 from comfy.ldm.flux.layers import timestep_embedding
 
+def get_silence_latent(length, device):
+    head = torch.tensor([[[ 0.5707,  0.0982,  0.6909, -0.5658,  0.6266,  0.6996, -0.1365, -0.1291,
+                        -0.0776, -0.1171, -0.2743, -0.8422, -0.1168,  1.5539, -4.6936,  0.7436,
+                        -1.1846, -0.2637,  0.6933, -6.7266,  0.0966, -0.1187, -0.3501, -1.1736,
+                        0.0587, -2.0517, -1.3651,  0.7508, -0.2490, -1.3548, -0.1290, -0.7261,
+                        1.1132, -0.3249,  0.2337,  0.3004,  0.6605, -0.0298, -0.1989, -0.4041,
+                        0.2843, -1.0963, -0.5519,  0.2639, -1.0436, -0.1183,  0.0640,  0.4460,
+                        -1.1001, -0.6172, -1.3241,  1.1379,  0.5623, -0.1507, -0.1963, -0.4742,
+                        -2.4697,  0.5302,  0.5381,  0.4636, -0.1782, -0.0687,  1.0333,  0.4202],
+                        [ 0.3040, -0.1367,  0.6200,  0.0665, -0.0642,  0.4655, -0.1187, -0.0440,
+                        0.2941, -0.2753,  0.0173, -0.2421, -0.0147,  1.5603, -2.7025,  0.7907,
+                        -0.9736, -0.0682,  0.1294, -5.0707, -0.2167,  0.3302, -0.1513, -0.8100,
+                        -0.3894, -0.2884, -0.3149,  0.8660, -0.3817, -1.7061,  0.5824, -0.4840,
+                        0.6938,  0.1859,  0.1753,  0.3081,  0.0195,  0.1403, -0.0754, -0.2091,
+                        0.1251, -0.1578, -0.4968, -0.1052, -0.4554, -0.0320,  0.1284,  0.4974,
+                        -1.1889, -0.0344, -0.8313,  0.2953,  0.5445, -0.6249, -0.1595, -0.0682,
+                        -3.1412,  0.0484,  0.4153,  0.8260, -0.1526, -0.0625,  0.5366,  0.8473],
+                        [ 5.3524e-02, -1.7534e-01,  5.4443e-01, -4.3501e-01, -2.1317e-03,
+                        3.7200e-01, -4.0143e-03, -1.5516e-01, -1.2968e-01, -1.5375e-01,
+                        -7.7107e-02, -2.0593e-01, -3.2780e-01,  1.5142e+00, -2.6101e+00,
+                        5.8698e-01, -1.2716e+00, -2.4773e-01, -2.7933e-02, -5.0799e+00,
+                        1.1601e-01,  4.0987e-01, -2.2030e-02, -6.6495e-01, -2.0995e-01,
+                        -6.3474e-01, -1.5893e-01,  8.2745e-01, -2.2992e-01, -1.6816e+00,
+                        5.4440e-01, -4.9579e-01,  5.5128e-01,  3.0477e-01,  8.3052e-02,
+                        -6.1782e-02,  5.9036e-03,  2.9553e-01, -8.0645e-02, -1.0060e-01,
+                        1.9144e-01, -3.8124e-01, -7.2949e-01,  2.4520e-02, -5.0814e-01,
+                        2.3977e-01,  9.2943e-02,  3.9256e-01, -1.1993e+00, -3.2752e-01,
+                        -7.2707e-01,  2.9476e-01,  4.3542e-01, -8.8597e-01, -4.1686e-01,
+                        -8.5390e-02, -2.9018e+00,  6.4988e-02,  5.3945e-01,  9.1988e-01,
+                        5.8762e-02, -7.0098e-02,  6.4772e-01,  8.9118e-01],
+                        [-3.2225e-02, -1.3195e-01,  5.6411e-01, -5.4766e-01, -5.2170e-03,
+                        3.1425e-01, -5.4367e-02, -1.9419e-01, -1.3059e-01, -1.3660e-01,
+                        -9.0984e-02, -1.9540e-01, -2.5590e-01,  1.5440e+00, -2.6349e+00,
+                        6.8273e-01, -1.2532e+00, -1.9810e-01, -2.2793e-02, -5.0506e+00,
+                        1.8818e-01,  5.0109e-01,  7.3546e-03, -6.8771e-01, -3.0676e-01,
+                        -7.3257e-01, -1.6687e-01,  9.2232e-01, -1.8987e-01, -1.7267e+00,
+                        5.3355e-01, -5.3179e-01,  4.4953e-01,  2.8820e-01,  1.3012e-01,
+                        -2.0943e-01, -1.1348e-01,  3.3929e-01, -1.5069e-01, -1.2919e-01,
+                        1.8929e-01, -3.6166e-01, -8.0756e-01,  6.6387e-02, -5.8867e-01,
+                        1.6978e-01,  1.0134e-01,  3.3877e-01, -1.2133e+00, -3.2492e-01,
+                        -8.1237e-01,  3.8101e-01,  4.3765e-01, -8.0596e-01, -4.4531e-01,
+                        -4.7513e-02, -2.9266e+00,  1.1741e-03,  4.5123e-01,  9.3075e-01,
+                        5.3688e-02, -1.9621e-01,  6.4530e-01,  9.3870e-01]]], device=device).movedim(-1, 1)
+
+    silence_latent = torch.tensor([[[-1.3672e-01, -1.5820e-01,  5.8594e-01, -5.7422e-01,  3.0273e-02,
+                                2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01,
+                                -2.7710e-02, -1.8066e-01, -2.9688e-01,  1.6016e+00, -2.6719e+00,
+                                7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00,
+                                2.4316e-01,  4.7266e-01,  4.6387e-02, -6.6406e-01, -2.1973e-01,
+                                -6.7578e-01, -1.5723e-01,  9.5312e-01, -2.0020e-01, -1.7109e+00,
+                                5.8984e-01, -5.7422e-01,  5.1562e-01,  2.8320e-01,  1.4551e-01,
+                                -1.8750e-01, -5.9814e-02,  3.6719e-01, -1.0059e-01, -1.5723e-01,
+                                2.0605e-01, -4.3359e-01, -8.2812e-01,  4.5654e-02, -6.6016e-01,
+                                1.4844e-01,  9.4727e-02,  3.8477e-01, -1.2578e+00, -3.3203e-01,
+                                -8.5547e-01,  4.3359e-01,  4.2383e-01, -8.9453e-01, -5.0391e-01,
+                                -5.6152e-02, -2.9219e+00, -2.4658e-02,  5.0391e-01,  9.8438e-01,
+                                7.2754e-02, -2.1582e-01,  6.3672e-01,  1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, length)
+    silence_latent[:, :, :head.shape[-1]] = head
+    return silence_latent
+
+
 def get_layer_class(operations, layer_name):
     if operations is not None and hasattr(operations, layer_name):
         return getattr(operations, layer_name)
@@ -183,7 +244,7 @@ class AceStepAttention(nn.Module):
             else:
                 attn_bias = window_bias
 
-        attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True)
+        attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True, low_precision_attention=False)
         attn_output = self.o_proj(attn_output)
 
         return attn_output
@@ -677,7 +738,7 @@ class AttentionPooler(nn.Module):
     def forward(self, x):
         B, T, P, D = x.shape
         x = self.embed_tokens(x)
-        special = self.special_token.expand(B, T, 1, -1)
+        special = comfy.model_management.cast_to(self.special_token, device=x.device, dtype=x.dtype).expand(B, T, 1, -1)
         x = torch.cat([special, x], dim=2)
         x = x.view(B * T, P + 1, D)
 
@@ -728,7 +789,7 @@ class FSQ(nn.Module):
         self.register_buffer('implicit_codebook', implicit_codebook, persistent=False)
 
     def bound(self, z):
-        levels_minus_1 = (self._levels - 1).to(z.dtype)
+        levels_minus_1 = (comfy.model_management.cast_to(self._levels, device=z.device, dtype=z.dtype) - 1)
         scale = 2. / levels_minus_1
         bracket = (levels_minus_1 * (torch.tanh(z) + 1) / 2.) + 0.5
 
@@ -743,8 +804,8 @@ class FSQ(nn.Module):
         return codes_non_centered.float() * (2. / (self._levels.float() - 1)) - 1.
 
     def codes_to_indices(self, zhat):
-        zhat_normalized = (zhat + 1.) / (2. / (self._levels.to(zhat.dtype) - 1))
-        return (zhat_normalized * self._basis.to(zhat.dtype)).sum(dim=-1).round().to(torch.int32)
+        zhat_normalized = (zhat + 1.) / (2. / (comfy.model_management.cast_to(self._levels, device=zhat.device, dtype=zhat.dtype) - 1))
+        return (zhat_normalized * comfy.model_management.cast_to(self._basis, device=zhat.device, dtype=zhat.dtype)).sum(dim=-1).round().to(torch.int32)
 
     def forward(self, z):
         orig_dtype = z.dtype
@@ -826,7 +887,7 @@ class ResidualFSQ(nn.Module):
         x = self.project_in(x)
 
         if hasattr(self, 'soft_clamp_input_value'):
-            sc_val = self.soft_clamp_input_value.to(x.dtype)
+            sc_val = comfy.model_management.cast_to(self.soft_clamp_input_value, device=x.device, dtype=x.dtype)
             x = (x / sc_val).tanh() * sc_val
 
         quantized_out = torch.tensor(0., device=x.device, dtype=x.dtype)
@@ -834,7 +895,7 @@ class ResidualFSQ(nn.Module):
         all_indices = []
 
         for layer, scale in zip(self.layers, self.scales):
-            scale = scale.to(residual.dtype)
+            scale = comfy.model_management.cast_to(scale, device=x.device, dtype=x.dtype)
 
             quantized, indices = layer(residual / scale)
             quantized = quantized * scale
@@ -1035,28 +1096,26 @@ class AceStepConditionGenerationModel(nn.Module):
                     audio_codes = torch.nn.functional.pad(audio_codes, (0, math.ceil(src_latents.shape[1] / 5) - audio_codes.shape[1]), "constant", 35847)
                 lm_hints_5Hz = self.tokenizer.quantizer.get_output_from_indices(audio_codes, dtype=text_hidden_states.dtype)
             else:
-                assert False
-                # TODO ?
+                lm_hints_5Hz, indices = self.tokenizer.tokenize(refer_audio_acoustic_hidden_states_packed)
 
             lm_hints = self.detokenizer(lm_hints_5Hz)
 
         lm_hints = lm_hints[:, :src_latents.shape[1], :]
-        if is_covers is None:
+        if is_covers is None or is_covers is True:
             src_latents = lm_hints
-        else:
-            src_latents = torch.where(is_covers.unsqueeze(-1).unsqueeze(-1) > 0, lm_hints, src_latents)
+        elif is_covers is False:
+            src_latents = refer_audio_acoustic_hidden_states_packed
 
         context_latents = torch.cat([src_latents, chunk_masks.to(src_latents.dtype)], dim=-1)
 
         return encoder_hidden, encoder_mask, context_latents
 
-    def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, **kwargs):
+    def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, is_covers=None, replace_with_null_embeds=False, **kwargs):
         text_attention_mask = None
         lyric_attention_mask = None
         refer_audio_order_mask = None
         attention_mask = None
         chunk_masks = None
-        is_covers = None
         src_latents = None
         precomputed_lm_hints_25Hz = None
         lyric_hidden_states = lyric_embed
@@ -1068,7 +1127,7 @@ class AceStepConditionGenerationModel(nn.Module):
         if refer_audio_order_mask is None:
             refer_audio_order_mask = torch.zeros((x.shape[0],), device=x.device, dtype=torch.long)
 
-        if src_latents is None and is_covers is None:
+        if src_latents is None:
             src_latents = x
 
         if chunk_masks is None:
@@ -1081,6 +1140,9 @@ class AceStepConditionGenerationModel(nn.Module):
             src_latents, chunk_masks, is_covers, precomputed_lm_hints_25Hz=precomputed_lm_hints_25Hz, audio_codes=audio_codes
         )
 
+        if replace_with_null_embeds:
+            enc_hidden[:] = self.null_condition_emb.to(enc_hidden)
+
         out = self.decoder(hidden_states=x,
                            timestep=timestep,
                            timestep_r=timestep,

comfy/ldm/anima/model.py

@@ -195,8 +195,20 @@ class Anima(MiniTrainDIT):
         super().__init__(*args, **kwargs)
         self.llm_adapter = LLMAdapter(device=kwargs.get("device"), dtype=kwargs.get("dtype"), operations=kwargs.get("operations"))
 
-    def preprocess_text_embeds(self, text_embeds, text_ids):
+    def preprocess_text_embeds(self, text_embeds, text_ids, t5xxl_weights=None):
         if text_ids is not None:
-            return self.llm_adapter(text_embeds, text_ids)
+            out = self.llm_adapter(text_embeds, text_ids)
+            if t5xxl_weights is not None:
+                out = out * t5xxl_weights
+
+            if out.shape[1] < 512:
+                out = torch.nn.functional.pad(out, (0, 0, 0, 512 - out.shape[1]))
+            return out
         else:
             return text_embeds
+
+    def forward(self, x, timesteps, context, **kwargs):
+        t5xxl_ids = kwargs.pop("t5xxl_ids", None)
+        if t5xxl_ids is not None:
+            context = self.preprocess_text_embeds(context, t5xxl_ids, t5xxl_weights=kwargs.pop("t5xxl_weights", None))
+        return super().forward(x, timesteps, context, **kwargs)

comfy/ldm/cosmos/predict2.py

@@ -335,7 +335,7 @@ class FinalLayer(nn.Module):
         device=None, dtype=None, operations=None
     ):
         super().__init__()
-        self.layer_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.layer_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
         self.linear = operations.Linear(
             hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, device=device, dtype=dtype
         )
@@ -463,6 +463,8 @@ class Block(nn.Module):
         extra_per_block_pos_emb: Optional[torch.Tensor] = None,
         transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
+        residual_dtype = x_B_T_H_W_D.dtype
+        compute_dtype = emb_B_T_D.dtype
         if extra_per_block_pos_emb is not None:
             x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
 
@@ -512,7 +514,7 @@ class Block(nn.Module):
         result_B_T_H_W_D = rearrange(
             self.self_attn(
                 # normalized_x_B_T_HW_D,
-                rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                rearrange(normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
                 None,
                 rope_emb=rope_emb_L_1_1_D,
                 transformer_options=transformer_options,
@@ -522,7 +524,7 @@ class Block(nn.Module):
             h=H,
             w=W,
         )
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D * result_B_T_H_W_D
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
 
         def _x_fn(
             _x_B_T_H_W_D: torch.Tensor,
@@ -536,7 +538,7 @@ class Block(nn.Module):
             )
             _result_B_T_H_W_D = rearrange(
                 self.cross_attn(
-                    rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                    rearrange(_normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
                     crossattn_emb,
                     rope_emb=rope_emb_L_1_1_D,
                     transformer_options=transformer_options,
@@ -555,7 +557,7 @@ class Block(nn.Module):
             shift_cross_attn_B_T_1_1_D,
             transformer_options=transformer_options,
         )
-        x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
+        x_B_T_H_W_D = result_B_T_H_W_D.to(residual_dtype) * gate_cross_attn_B_T_1_1_D.to(residual_dtype) + x_B_T_H_W_D
 
         normalized_x_B_T_H_W_D = _fn(
             x_B_T_H_W_D,
@@ -563,8 +565,8 @@ class Block(nn.Module):
             scale_mlp_B_T_1_1_D,
             shift_mlp_B_T_1_1_D,
         )
-        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D)
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D * result_B_T_H_W_D
+        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D.to(compute_dtype))
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
         return x_B_T_H_W_D
 
 
@@ -876,6 +878,14 @@ class MiniTrainDIT(nn.Module):
             "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
             "transformer_options": kwargs.get("transformer_options", {}),
         }
+
+        # The residual stream for this model has large values. To make fp16 compute_dtype work, we keep the residual stream
+        # in fp32, but run attention and MLP modules in fp16.
+        # An alternate method that clamps fp16 values "works" in the sense that it makes coherent images, but there is noticeable
+        # quality degradation and visual artifacts.
+        if x_B_T_H_W_D.dtype == torch.float16:
+            x_B_T_H_W_D = x_B_T_H_W_D.float()
+
         for block in self.blocks:
             x_B_T_H_W_D = block(
                 x_B_T_H_W_D,
@@ -884,6 +894,6 @@ class MiniTrainDIT(nn.Module):
                 **block_kwargs,
             )
 
-        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
+        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D.to(crossattn_emb.dtype), t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
         x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)[:, :, :orig_shape[-3], :orig_shape[-2], :orig_shape[-1]]
         return x_B_C_Tt_Hp_Wp

comfy/ldm/flux/math.py

@@ -29,19 +29,34 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     return out.to(dtype=torch.float32, device=pos.device)
 
 
+def _apply_rope1(x: Tensor, freqs_cis: Tensor):
+    x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
+
+    x_out = freqs_cis[..., 0] * x_[..., 0]
+    x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
+
+    return x_out.reshape(*x.shape).type_as(x)
+
+
+def _apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+    return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+
+
 try:
     import comfy.quant_ops
-    apply_rope = comfy.quant_ops.ck.apply_rope
-    apply_rope1 = comfy.quant_ops.ck.apply_rope1
+    q_apply_rope = comfy.quant_ops.ck.apply_rope
+    q_apply_rope1 = comfy.quant_ops.ck.apply_rope1
+    def apply_rope(xq, xk, freqs_cis):
+        if comfy.model_management.in_training:
+            return _apply_rope(xq, xk, freqs_cis)
+        else:
+            return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+    def apply_rope1(x, freqs_cis):
+        if comfy.model_management.in_training:
+            return _apply_rope1(x, freqs_cis)
+        else:
+            return q_apply_rope1(x, freqs_cis)
 except:
     logging.warning("No comfy kitchen, using old apply_rope functions.")
-    def apply_rope1(x: Tensor, freqs_cis: Tensor):
-        x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
-
-        x_out = freqs_cis[..., 0] * x_[..., 0]
-        x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
-
-        return x_out.reshape(*x.shape).type_as(x)
-
-    def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-        return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+    apply_rope = _apply_rope
+    apply_rope1 = _apply_rope1

comfy/ldm/modules/attention.py

@@ -524,6 +524,9 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
 
 @wrap_attn
 def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    if kwargs.get("low_precision_attention", True) is False:
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=skip_reshape, skip_output_reshape=skip_output_reshape, **kwargs)
+
     exception_fallback = False
     if skip_reshape:
         b, _, _, dim_head = q.shape

comfy/lora.py

@@ -332,6 +332,12 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["{}".format(key_lora)] = k
                 key_map["transformer.{}".format(key_lora)] = k
 
+    if isinstance(model, comfy.model_base.ACEStep15):
+        for k in sdk:
+            if k.startswith("diffusion_model.decoder.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model.decoder."):-len(".weight")]
+                key_map["base_model.model.{}".format(key_lora)] = k  # Official base model loras
+
     return key_map
 
 

comfy/model_base.py

@@ -147,11 +147,11 @@ class BaseModel(torch.nn.Module):
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
             logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
+            comfy.model_management.archive_model_dtypes(self.diffusion_model)
+
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
-        comfy.model_management.archive_model_dtypes(self.diffusion_model)
-
         self.adm_channels = unet_config.get("adm_in_channels", None)
         if self.adm_channels is None:
             self.adm_channels = 0
@@ -1160,12 +1160,16 @@ class Anima(BaseModel):
         device = kwargs["device"]
         if cross_attn is not None:
             if t5xxl_ids is not None:
-                cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.unsqueeze(0).to(device=device))
                 if t5xxl_weights is not None:
-                    cross_attn *= t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn)
+                    t5xxl_weights = t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn)
+                t5xxl_ids = t5xxl_ids.unsqueeze(0)
+
+                if torch.is_inference_mode_enabled():  # if not we are training
+                    cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.to(device=device), t5xxl_weights=t5xxl_weights.to(device=device, dtype=self.get_dtype()))
+                else:
+                    out['t5xxl_ids'] = comfy.conds.CONDRegular(t5xxl_ids)
+                    out['t5xxl_weights'] = comfy.conds.CONDRegular(t5xxl_weights)
 
-                if cross_attn.shape[1] < 512:
-                    cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, 0, 512 - cross_attn.shape[1]))
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
@@ -1548,9 +1552,12 @@ class ACEStep15(BaseModel):
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
         device = kwargs["device"]
+        noise = kwargs["noise"]
 
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
+            if torch.count_nonzero(cross_attn) == 0:
+                out['replace_with_null_embeds'] = comfy.conds.CONDConstant(True)
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
 
         conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
@@ -1559,27 +1566,26 @@ class ACEStep15(BaseModel):
 
         refer_audio = kwargs.get("reference_audio_timbre_latents", None)
         if refer_audio is None or len(refer_audio) == 0:
-            refer_audio = torch.tensor([[[-1.3672e-01, -1.5820e-01,  5.8594e-01, -5.7422e-01,  3.0273e-02,
-                                        2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01,
-                                        -2.7710e-02, -1.8066e-01, -2.9688e-01,  1.6016e+00, -2.6719e+00,
-                                        7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00,
-                                        2.4316e-01,  4.7266e-01,  4.6387e-02, -6.6406e-01, -2.1973e-01,
-                                        -6.7578e-01, -1.5723e-01,  9.5312e-01, -2.0020e-01, -1.7109e+00,
-                                        5.8984e-01, -5.7422e-01,  5.1562e-01,  2.8320e-01,  1.4551e-01,
-                                        -1.8750e-01, -5.9814e-02,  3.6719e-01, -1.0059e-01, -1.5723e-01,
-                                        2.0605e-01, -4.3359e-01, -8.2812e-01,  4.5654e-02, -6.6016e-01,
-                                        1.4844e-01,  9.4727e-02,  3.8477e-01, -1.2578e+00, -3.3203e-01,
-                                        -8.5547e-01,  4.3359e-01,  4.2383e-01, -8.9453e-01, -5.0391e-01,
-                                        -5.6152e-02, -2.9219e+00, -2.4658e-02,  5.0391e-01,  9.8438e-01,
-                                        7.2754e-02, -2.1582e-01,  6.3672e-01,  1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, 750)
+            refer_audio = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            pass_audio_codes = True
         else:
-            refer_audio = refer_audio[-1]
+            refer_audio = refer_audio[-1][:, :, :noise.shape[2]]
+            out['is_covers'] = comfy.conds.CONDConstant(True)
+            pass_audio_codes = False
+
+        if pass_audio_codes:
+            audio_codes = kwargs.get("audio_codes", None)
+            if audio_codes is not None:
+                out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
+                refer_audio = refer_audio[:, :, :750]
+            else:
+                out['is_covers'] = comfy.conds.CONDConstant(False)
+
+        if refer_audio.shape[2] < noise.shape[2]:
+            pad = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            refer_audio = torch.cat([refer_audio.to(pad), pad[:, :, refer_audio.shape[2]:]], dim=2)
+
         out['refer_audio'] = comfy.conds.CONDRegular(refer_audio)
-
-        audio_codes = kwargs.get("audio_codes", None)
-        if audio_codes is not None:
-            out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
-
         return out
 
 class Omnigen2(BaseModel):

comfy/model_management.py

@@ -19,7 +19,7 @@
 import psutil
 import logging
 from enum import Enum
-from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
+from comfy.cli_args import args, PerformanceFeature
 import threading
 import torch
 import sys
@@ -55,6 +55,11 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
+
+# Training Related State
+in_training = False
+
+
 def get_supported_float8_types():
     float8_types = []
     try:
@@ -651,7 +656,7 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
                 soft_empty_cache()
     return unloaded_models
 
-def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     cleanup_models_gc()
     global vram_state
 
@@ -747,26 +752,6 @@ def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, m
         current_loaded_models.insert(0, loaded_model)
     return
 
-def load_models_gpu_thread(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load):
-    with torch.inference_mode():
-        load_models_gpu_orig(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
-        soft_empty_cache()
-
-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
-    #Deliberately load models outside of the Aimdo mempool so they can be retained accross
-    #nodes. Use a dummy thread to do it as pytorch documents that mempool contexts are
-    #thread local. So exploit that to escape context
-    if enables_dynamic_vram():
-        t = threading.Thread(
-            target=load_models_gpu_thread,
-            args=(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
-        )
-        t.start()
-        t.join()
-    else:
-        load_models_gpu_orig(models, memory_required=memory_required, force_patch_weights=force_patch_weights,
-                             minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
-
 def load_model_gpu(model):
     return load_models_gpu([model])
 
@@ -1226,21 +1211,16 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
         if dtype is None:
             dtype = weight._model_dtype
 
-        r = torch.empty_like(weight, dtype=dtype, device=device)
-
         signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
         if signature is not None:
-            raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
-            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0]
+            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, weight._v_tensor)[0]
             if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
                 weight._v_signature = signature
                 #Send it over
                 v_tensor.copy_(weight, non_blocking=non_blocking)
-            #always take a deep copy even if _v is good, as we have no reasonable point to unpin
-            #a non comfy weight
-            r.copy_(v_tensor)
-            comfy_aimdo.model_vbar.vbar_unpin(weight._v)
-            return r
+            return v_tensor.to(dtype=dtype)
+
+        r = torch.empty_like(weight, dtype=dtype, device=device)
 
         if weight.dtype != r.dtype and weight.dtype != weight._model_dtype:
             #Offloaded casting could skip this, however it would make the quantizations
@@ -1724,11 +1704,9 @@ def soft_empty_cache(force=False):
     elif is_mlu():
         torch.mlu.empty_cache()
     elif torch.cuda.is_available():
-        if comfy.memory_management.aimdo_allocator is None:
-            #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist
-            torch.cuda.synchronize()
-            torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
+        torch.cuda.synchronize()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
 
 def unload_all_models():
     free_memory(1e30, get_torch_device())

comfy/model_patcher.py

@@ -19,7 +19,6 @@
 from __future__ import annotations
 
 import collections
-import copy
 import inspect
 import logging
 import math
@@ -317,7 +316,7 @@ class ModelPatcher:
 
         n.object_patches = self.object_patches.copy()
         n.weight_wrapper_patches = self.weight_wrapper_patches.copy()
-        n.model_options = copy.deepcopy(self.model_options)
+        n.model_options = comfy.utils.deepcopy_list_dict(self.model_options)
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
@@ -1400,7 +1399,7 @@ class ModelPatcher:
                 continue
             key = "diffusion_model." + k
             unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key))
-        return self.model.state_dict_for_saving(unet_state_dict)
+        return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
 
     def __del__(self):
         self.unpin_all_weights()
@@ -1492,7 +1491,9 @@ class ModelPatcherDynamic(ModelPatcher):
             if vbar is not None:
                 vbar.prioritize()
 
-            #We have way more tools for acceleration on comfy weight offloading, so always
+            #We force reserve VRAM for the non comfy-weight so we dont have to deal
+            #with pin and unpin syncrhonization which can be expensive for small weights
+            #with a high layer rate (e.g. autoregressive LLMs).
             #prioritize the non-comfy weights (note the order reverse).
             loading = self._load_list(prio_comfy_cast_weights=True)
             loading.sort(reverse=True)
@@ -1541,6 +1542,7 @@ class ModelPatcherDynamic(ModelPatcher):
 
                     if vbar is not None and not hasattr(m, "_v"):
                         m._v = vbar.alloc(v_weight_size)
+                        m._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(m._v, device_to)
                     allocated_size += v_weight_size
 
                 else:
@@ -1555,8 +1557,10 @@ class ModelPatcherDynamic(ModelPatcher):
                         weight_size = geometry.numel() * geometry.element_size()
                         if vbar is not None and not hasattr(weight, "_v"):
                             weight._v = vbar.alloc(weight_size)
+                            weight._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device_to)
                             weight._model_dtype = model_dtype
                         allocated_size += weight_size
+                    vbar.set_watermark_limit(allocated_size)
 
             logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.")
 

comfy/ops.py

@@ -54,6 +54,8 @@ try:
             SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
 
             def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+                if q.nelement() < 1024 * 128:  # arbitrary number, for small inputs cudnn attention seems slower
+                    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
                 with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
                     return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
         else:
@@ -85,7 +87,7 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
 
     signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
     if signature is not None:
-        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+        xfer_dest = s._v_tensor
     resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
 
     if not resident:
@@ -167,8 +169,8 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
                 if orig.dtype == dtype and len(fns) == 0:
                     #The layer actually wants our freshly saved QT
                     x = y
-            else:
-                y = x
+            elif update_weight:
+                y = comfy.float.stochastic_rounding(x, orig.dtype, seed = comfy.utils.string_to_seed(s.seed_key))
             if update_weight:
                 orig.copy_(y)
         for f in fns:

comfy/sampler_helpers.py

@@ -122,20 +122,26 @@ def estimate_memory(model, noise_shape, conds):
     minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min)
     return memory_required, minimum_memory_required
 
-def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
+def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _prepare_sampling,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True)
     )
-    return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load)
+    return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load, force_offload=force_offload)
 
-def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
+def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False):
     real_model: BaseModel = None
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
-    memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
-    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory, force_full_load=force_full_load)
+    if force_offload: # In training + offload enabled, we want to force prepare sampling to trigger partial load
+        memory_required = 1e20
+        minimum_memory_required = None
+    else:
+        memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
+        memory_required += inference_memory
+        minimum_memory_required += inference_memory
+    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
     real_model = model.model
 
     return real_model, conds, models

comfy/sd.py

@@ -554,6 +554,8 @@ class VAE:
             elif "decoder.layers.1.layers.0.beta" in sd:
                 config = {}
                 param_key = None
+                self.upscale_ratio = 2048
+                self.downscale_ratio = 2048
                 if "decoder.layers.2.layers.1.weight_v" in sd:
                     param_key = "decoder.layers.2.layers.1.weight_v"
                 if "decoder.layers.2.layers.1.parametrizations.weight.original1" in sd:
@@ -562,6 +564,8 @@ class VAE:
                     if sd[param_key].shape[-1] == 12:
                         config["strides"] = [2, 4, 4, 6, 10]
                         self.audio_sample_rate = 48000
+                        self.upscale_ratio = 1920
+                        self.downscale_ratio = 1920
 
                 self.first_stage_model = AudioOobleckVAE(**config)
                 self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)
@@ -569,8 +573,6 @@ class VAE:
                 self.latent_channels = 64
                 self.output_channels = 2
                 self.pad_channel_value = "replicate"
-                self.upscale_ratio = 2048
-                self.downscale_ratio = 2048
                 self.latent_dim = 1
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
@@ -791,8 +793,6 @@ class VAE:
             self.first_stage_model = AutoencoderKL(**(config['params']))
         self.first_stage_model = self.first_stage_model.eval()
 
-        model_management.archive_model_dtypes(self.first_stage_model)
-
         if device is None:
             device = model_management.vae_device()
         self.device = device
@@ -801,6 +801,7 @@ class VAE:
             dtype = model_management.vae_dtype(self.device, self.working_dtypes)
         self.vae_dtype = dtype
         self.first_stage_model.to(self.vae_dtype)
+        model_management.archive_model_dtypes(self.first_stage_model)
         self.output_device = model_management.intermediate_device()
 
         mp = comfy.model_patcher.CoreModelPatcher
@@ -870,7 +871,7 @@ class VAE:
             / 3.0)
         return output
 
-    def decode_tiled_1d(self, samples, tile_x=128, overlap=32):
+    def decode_tiled_1d(self, samples, tile_x=256, overlap=32):
         if samples.ndim == 3:
             decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
         else:
@@ -974,7 +975,7 @@ class VAE:
         if overlap is not None:
             args["overlap"] = overlap
 
-        if dims == 1:
+        if dims == 1 or self.extra_1d_channel is not None:
             args.pop("tile_y")
             output = self.decode_tiled_1d(samples, **args)
         elif dims == 2:
@@ -1442,7 +1443,12 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
             tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
         elif clip_type == CLIPType.ACE:
-            clip_target.clip = comfy.text_encoders.ace15.te(**llama_detect(clip_data))
+            te_models = [detect_te_model(clip_data[0]), detect_te_model(clip_data[1])]
+            if TEModel.QWEN3_4B in te_models:
+                model_type = "qwen3_4b"
+            else:
+                model_type = "qwen3_2b"
+            clip_target.clip = comfy.text_encoders.ace15.te(lm_model=model_type, **llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel

comfy/supported_models.py

@@ -993,7 +993,7 @@ class CosmosT2IPredict2(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
@@ -1023,11 +1023,7 @@ class Anima(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
-
-    def __init__(self, unet_config):
-        super().__init__(unet_config)
-        self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.Anima(self, device=device)
@@ -1038,6 +1034,12 @@ class Anima(supported_models_base.BASE):
         detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_06b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.anima.AnimaTokenizer, comfy.text_encoders.anima.te(**detect))
 
+    def set_inference_dtype(self, dtype, manual_cast_dtype, **kwargs):
+        self.memory_usage_factor = (self.unet_config.get("model_channels", 2048) / 2048) * 0.95
+        if dtype is torch.float16:
+            self.memory_usage_factor *= 1.4
+        return super().set_inference_dtype(dtype, manual_cast_dtype, **kwargs)
+
 class CosmosI2VPredict2(CosmosT2IPredict2):
     unet_config = {
         "image_model": "cosmos_predict2",
@@ -1625,8 +1627,16 @@ class ACEStep15(supported_models_base.BASE):
 
     def clip_target(self, state_dict={}):
         pref = self.text_encoder_key_prefix[0]
-        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**hunyuan_detect))
+        detect_2b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
+        detect_4b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref))
+        if "dtype_llama" in detect_2b:
+            detect = detect_2b
+            detect["lm_model"] = "qwen3_2b"
+        elif "dtype_llama" in detect_4b:
+            detect = detect_4b
+            detect["lm_model"] = "qwen3_4b"
+
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**detect))
 
 
 models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]

comfy/text_encoders/ace15.py

@@ -3,6 +3,9 @@ import comfy.text_encoders.llama
 from comfy import sd1_clip
 import torch
 import math
+from tqdm.auto import trange
+import yaml
+import comfy.utils
 
 
 def sample_manual_loop_no_classes(
@@ -18,8 +21,11 @@ def sample_manual_loop_no_classes(
     min_tokens: int = 1,
     max_new_tokens: int = 2048,
     audio_start_id: int = 151669,  # The cutoff ID for audio codes
+    audio_end_id: int = 215669,
     eos_token_id: int = 151645,
 ):
+    if ids is None:
+        return []
     device = model.execution_device
 
     if execution_dtype is None:
@@ -29,6 +35,7 @@ def sample_manual_loop_no_classes(
             execution_dtype = torch.float32
 
     embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device)
+    embeds_batch = embeds.shape[0]
     for i, t in enumerate(paddings):
         attention_mask[i, :t] = 0
         attention_mask[i, t:] = 1
@@ -38,32 +45,41 @@ def sample_manual_loop_no_classes(
     generator = torch.Generator(device=device)
     generator.manual_seed(seed)
     model_config = model.transformer.model.config
+    past_kv_shape = [embeds_batch, model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim]
 
     for x in range(model_config.num_hidden_layers):
-        past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), 0))
+        past_key_values.append((torch.empty(past_kv_shape, device=device, dtype=execution_dtype), torch.empty(past_kv_shape, device=device, dtype=execution_dtype), 0))
 
-    for step in range(max_new_tokens):
+    progress_bar = comfy.utils.ProgressBar(max_new_tokens)
+
+    for step in trange(max_new_tokens, desc="LM sampling"):
         outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values)
         next_token_logits = model.transformer.logits(outputs[0])[:, -1]
         past_key_values = outputs[2]
 
-        cond_logits = next_token_logits[0:1]
-        uncond_logits = next_token_logits[1:2]
-        cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits)
+        if cfg_scale != 1.0:
+            cond_logits = next_token_logits[0:1]
+            uncond_logits = next_token_logits[1:2]
+            cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits)
+        else:
+            cfg_logits = next_token_logits[0:1]
 
-        if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
+        use_eos_score = eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step
+        if use_eos_score:
             eos_score = cfg_logits[:, eos_token_id].clone()
 
+        remove_logit_value = torch.finfo(cfg_logits.dtype).min
         # Only generate audio tokens
-        cfg_logits[:, :audio_start_id] = float('-inf')
+        cfg_logits[:, :audio_start_id] = remove_logit_value
+        cfg_logits[:, audio_end_id:] = remove_logit_value
 
-        if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
+        if use_eos_score:
             cfg_logits[:, eos_token_id] = eos_score
 
         if top_k is not None and top_k > 0:
             top_k_vals, _ = torch.topk(cfg_logits, top_k)
             min_val = top_k_vals[..., -1, None]
-            cfg_logits[cfg_logits < min_val] = float('-inf')
+            cfg_logits[cfg_logits < min_val] = remove_logit_value
 
         if top_p is not None and top_p < 1.0:
             sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True)
@@ -72,7 +88,7 @@ def sample_manual_loop_no_classes(
             sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
             sorted_indices_to_remove[..., 0] = 0
             indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
-            cfg_logits[indices_to_remove] = float('-inf')
+            cfg_logits[indices_to_remove] = remove_logit_value
 
         if temperature > 0:
             cfg_logits = cfg_logits / temperature
@@ -86,61 +102,145 @@ def sample_manual_loop_no_classes(
             break
 
         embed, _, _, _ = model.process_tokens([[token]], device)
-        embeds = embed.repeat(2, 1, 1)
-        attention_mask = torch.cat([attention_mask, torch.ones((2, 1), device=device, dtype=attention_mask.dtype)], dim=1)
+        embeds = embed.repeat(embeds_batch, 1, 1)
+        attention_mask = torch.cat([attention_mask, torch.ones((embeds_batch, 1), device=device, dtype=attention_mask.dtype)], dim=1)
 
         output_audio_codes.append(token - audio_start_id)
+        progress_bar.update_absolute(step)
 
     return output_audio_codes
 
 
-def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0):
-    cfg_scale = 2.0
-
+def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0, cfg_scale=2.0, temperature=0.85, top_p=0.9, top_k=0):
     positive = [[token for token, _ in inner_list] for inner_list in positive]
-    negative = [[token for token, _ in inner_list] for inner_list in negative]
     positive = positive[0]
-    negative = negative[0]
 
-    neg_pad = 0
-    if len(negative) < len(positive):
-        neg_pad = (len(positive) - len(negative))
-        negative = [model.special_tokens["pad"]] * neg_pad + negative
+    if cfg_scale != 1.0:
+        negative = [[token for token, _ in inner_list] for inner_list in negative]
+        negative = negative[0]
 
-    pos_pad = 0
-    if len(negative) > len(positive):
-        pos_pad = (len(negative) - len(positive))
-        positive = [model.special_tokens["pad"]] * pos_pad + positive
+        neg_pad = 0
+        if len(negative) < len(positive):
+            neg_pad = (len(positive) - len(negative))
+            negative = [model.special_tokens["pad"]] * neg_pad + negative
 
-    paddings = [pos_pad, neg_pad]
-    return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
+        pos_pad = 0
+        if len(negative) > len(positive):
+            pos_pad = (len(negative) - len(positive))
+            positive = [model.special_tokens["pad"]] * pos_pad + positive
+
+        paddings = [pos_pad, neg_pad]
+        ids = [positive, negative]
+    else:
+        paddings = []
+        ids = [positive]
+
+    return sample_manual_loop_no_classes(model, ids, paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
 
 
 class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer)
 
+    def _metas_to_cot(self, *, return_yaml: bool = False, **kwargs) -> str:
+        user_metas = {
+            k: kwargs.pop(k)
+            for k in ("bpm", "duration", "keyscale", "timesignature")
+            if k in kwargs
+        }
+        timesignature = user_metas.get("timesignature")
+        if isinstance(timesignature, str) and timesignature.endswith("/4"):
+            user_metas["timesignature"] = timesignature[:-2]
+        user_metas = {
+            k: v if not isinstance(v, str) or not v.isdigit() else int(v)
+            for k, v in user_metas.items()
+            if v not in {"unspecified", None}
+        }
+        if len(user_metas):
+            meta_yaml = yaml.dump(user_metas, allow_unicode=True, sort_keys=True).strip()
+        else:
+            meta_yaml = ""
+        return f"<think>\n{meta_yaml}\n</think>" if not return_yaml else meta_yaml
+
+    def _metas_to_cap(self, **kwargs) -> str:
+        use_keys = ("bpm", "timesignature", "keyscale", "duration")
+        user_metas = { k: kwargs.pop(k, "N/A") for k in use_keys }
+        timesignature = user_metas.get("timesignature")
+        if isinstance(timesignature, str) and timesignature.endswith("/4"):
+            user_metas["timesignature"] = timesignature[:-2]
+        duration = user_metas["duration"]
+        if duration == "N/A":
+            user_metas["duration"] = "30 seconds"
+        elif isinstance(duration, (str, int, float)):
+            user_metas["duration"] = f"{math.ceil(float(duration))} seconds"
+        else:
+            raise TypeError("Unexpected type for duration key, must be str, int or float")
+        return "\n".join(f"- {k}: {user_metas[k]}" for k in use_keys)
+
     def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
-        out = {}
+        text = text.strip()
+        text_negative = kwargs.get("caption_negative", text).strip()
         lyrics = kwargs.get("lyrics", "")
-        bpm = kwargs.get("bpm", 120)
+        lyrics_negative = kwargs.get("lyrics_negative", lyrics)
         duration = kwargs.get("duration", 120)
-        keyscale = kwargs.get("keyscale", "C major")
-        timesignature = kwargs.get("timesignature", 2)
-        language = kwargs.get("language", "en")
+        if isinstance(duration, str):
+            duration = float(duration.split(None, 1)[0])
+        language = kwargs.get("language")
         seed = kwargs.get("seed", 0)
 
+        generate_audio_codes = kwargs.get("generate_audio_codes", True)
+        cfg_scale = kwargs.get("cfg_scale", 2.0)
+        temperature = kwargs.get("temperature", 0.85)
+        top_p = kwargs.get("top_p", 0.9)
+        top_k = kwargs.get("top_k", 0.0)
+
         duration = math.ceil(duration)
-        meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature)
-        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n<think>\n{}\n</think>\n\n<|im_end|>\n"
+        kwargs["duration"] = duration
+        tokens_duration = duration * 5
+        min_tokens = int(kwargs.get("min_tokens", tokens_duration))
+        max_tokens = int(kwargs.get("max_tokens", tokens_duration))
 
-        meta_cap = '- bpm: {}\n- timesignature: {}\n- keyscale: {}\n- duration: {}\n'.format(bpm, timesignature, keyscale, duration)
-        out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, meta_lm), disable_weights=True)
-        out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, ""), disable_weights=True)
+        metas_negative = {
+            k.rsplit("_", 1)[0]: kwargs.pop(k)
+            for k in ("bpm_negative", "duration_negative", "keyscale_negative", "timesignature_negative", "language_negative", "caption_negative")
+            if k in kwargs
+        }
+        if not kwargs.get("use_negative_caption"):
+            _ = metas_negative.pop("caption", None)
 
-        out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs)
-        out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs)
-        out["lm_metadata"] = {"min_tokens": duration * 5, "seed": seed}
+        cot_text = self._metas_to_cot(caption=text, **kwargs)
+        cot_text_negative = "<think>\n\n</think>" if not metas_negative else self._metas_to_cot(**metas_negative)
+        meta_cap = self._metas_to_cap(**kwargs)
+
+        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n\n# Lyric\n{}\n<|im_end|>\n<|im_start|>assistant\n{}\n\n<|im_end|>\n"
+        lyrics_template = "# Languages\n{}\n\n# Lyric\n{}<|endoftext|><|endoftext|>"
+        qwen3_06b_template = "# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}\n\n# Metas\n{}\n<|endoftext|>\n<|endoftext|>"
+
+        llm_prompts = {
+            "lm_prompt": lm_template.format(text, lyrics.strip(), cot_text),
+            "lm_prompt_negative": lm_template.format(text_negative, lyrics_negative.strip(), cot_text_negative),
+            "lyrics": lyrics_template.format(language if language is not None else "", lyrics),
+            "qwen3_06b": qwen3_06b_template.format(text, meta_cap),
+        }
+
+        out = {
+            prompt_key: self.qwen3_06b.tokenize_with_weights(
+                prompt,
+                prompt_key == "qwen3_06b" and return_word_ids,
+                disable_weights = True,
+                **kwargs,
+            )
+            for prompt_key, prompt in llm_prompts.items()
+        }
+        out["lm_metadata"] = {"min_tokens": min_tokens,
+                              "max_tokens": max_tokens,
+                              "seed": seed,
+                              "generate_audio_codes": generate_audio_codes,
+                              "cfg_scale": cfg_scale,
+                              "temperature": temperature,
+                              "top_p": top_p,
+                              "top_k": top_k,
+                              }
         return out
 
 
@@ -157,14 +257,34 @@ class Qwen3_2B_ACE15(sd1_clip.SDClipModel):
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_2B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
+class Qwen3_4B_ACE15(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
 class ACE15TEModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None, dtype_llama=None, model_options={}):
+    def __init__(self, device="cpu", dtype=None, dtype_llama=None, lm_model=None, model_options={}):
         super().__init__()
         if dtype_llama is None:
             dtype_llama = dtype
 
+        model = None
+        self.constant = 0.4375
+        if lm_model == "qwen3_4b":
+            model = Qwen3_4B_ACE15
+            self.constant = 0.5625
+        elif lm_model == "qwen3_2b":
+            model = Qwen3_2B_ACE15
+
+        self.lm_model = lm_model
         self.qwen3_06b = Qwen3_06BModel(device=device, dtype=dtype, model_options=model_options)
-        self.qwen3_2b = Qwen3_2B_ACE15(device=device, dtype=dtype_llama, model_options=model_options)
+        if model is not None:
+            setattr(self, self.lm_model, model(device=device, dtype=dtype_llama, model_options=model_options))
+
         self.dtypes = set([dtype, dtype_llama])
 
     def encode_token_weights(self, token_weight_pairs):
@@ -176,18 +296,26 @@ class ACE15TEModel(torch.nn.Module):
         self.qwen3_06b.set_clip_options({"layer": [0]})
         lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics)
 
-        lm_metadata = token_weight_pairs["lm_metadata"]
-        audio_codes = generate_audio_codes(self.qwen3_2b, token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"])
+        out = {"conditioning_lyrics": lyrics_embeds[:, 0]}
 
-        return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]}
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        if lm_metadata["generate_audio_codes"]:
+            audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["max_tokens"], seed=lm_metadata["seed"], cfg_scale=lm_metadata["cfg_scale"], temperature=lm_metadata["temperature"], top_p=lm_metadata["top_p"], top_k=lm_metadata["top_k"])
+            out["audio_codes"] = [audio_codes]
+
+        return base_out, None, out
 
     def set_clip_options(self, options):
         self.qwen3_06b.set_clip_options(options)
-        self.qwen3_2b.set_clip_options(options)
+        lm_model = getattr(self, self.lm_model, None)
+        if lm_model is not None:
+            lm_model.set_clip_options(options)
 
     def reset_clip_options(self):
         self.qwen3_06b.reset_clip_options()
-        self.qwen3_2b.reset_clip_options()
+        lm_model = getattr(self, self.lm_model, None)
+        if lm_model is not None:
+            lm_model.reset_clip_options()
 
     def load_sd(self, sd):
         if "model.layers.0.post_attention_layernorm.weight" in sd:
@@ -195,11 +323,11 @@ class ACE15TEModel(torch.nn.Module):
             if shape[0] == 1024:
                 return self.qwen3_06b.load_sd(sd)
             else:
-                return self.qwen3_2b.load_sd(sd)
+                return getattr(self, self.lm_model).load_sd(sd)
 
     def memory_estimation_function(self, token_weight_pairs, device=None):
         lm_metadata = token_weight_pairs["lm_metadata"]
-        constant = 0.4375
+        constant = self.constant
         if comfy.model_management.should_use_bf16(device):
             constant *= 0.5
 
@@ -208,11 +336,11 @@ class ACE15TEModel(torch.nn.Module):
         num_tokens += lm_metadata['min_tokens']
         return num_tokens * constant * 1024 * 1024
 
-def te(dtype_llama=None, llama_quantization_metadata=None):
+def te(dtype_llama=None, llama_quantization_metadata=None, lm_model="qwen3_2b"):
     class ACE15TEModel_(ACE15TEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
             if llama_quantization_metadata is not None:
                 model_options = model_options.copy()
                 model_options["llama_quantization_metadata"] = llama_quantization_metadata
-            super().__init__(device=device, dtype_llama=dtype_llama, dtype=dtype, model_options=model_options)
+            super().__init__(device=device, dtype_llama=dtype_llama, lm_model=lm_model, dtype=dtype, model_options=model_options)
     return ACE15TEModel_

comfy/text_encoders/anima.py

@@ -23,7 +23,7 @@ class AnimaTokenizer:
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
         qwen_ids = self.qwen3_06b.tokenize_with_weights(text, return_word_ids, **kwargs)
-        out["qwen3_06b"] = [[(token, 1.0) for token, _ in inner_list] for inner_list in qwen_ids]  # Set weights to 1.0
+        out["qwen3_06b"] = [[(k[0], 1.0, k[2]) if return_word_ids else (k[0], 1.0) for k in inner_list] for inner_list in qwen_ids]  # Set weights to 1.0
         out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 

comfy/text_encoders/llama.py

@@ -6,6 +6,7 @@ import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
+import comfy.ops
 import comfy.ldm.common_dit
 import comfy.clip_model
 
@@ -149,6 +150,29 @@ class Qwen3_2B_ACE15_lm_Config:
     final_norm: bool = True
     lm_head: bool = False
 
+@dataclass
+class Qwen3_4B_ACE15_lm_Config:
+    vocab_size: int = 217204
+    hidden_size: int = 2560
+    intermediate_size: int = 9728
+    num_hidden_layers: int = 36
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
 @dataclass
 class Qwen3_4BConfig:
     vocab_size: int = 151936
@@ -627,10 +651,10 @@ class Llama2_(nn.Module):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1])
-            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
+            mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min / 4)
 
         if seq_len > 1:
-            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
+            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min / 4).triu_(1)
             if mask is not None:
                 mask += causal_mask
             else:
@@ -738,6 +762,21 @@ class BaseLlama:
     def forward(self, input_ids, *args, **kwargs):
         return self.model(input_ids, *args, **kwargs)
 
+class BaseQwen3:
+    def logits(self, x):
+        input = x[:, -1:]
+        module = self.model.embed_tokens
+
+        offload_stream = None
+        if module.comfy_cast_weights:
+            weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True)
+        else:
+            weight = self.model.embed_tokens.weight.to(x)
+
+        x = torch.nn.functional.linear(input, weight, None)
+
+        comfy.ops.uncast_bias_weight(module, weight, None, offload_stream)
+        return x
 
 class Llama2(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
@@ -766,7 +805,7 @@ class Qwen25_3B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_06B(BaseLlama, torch.nn.Module):
+class Qwen3_06B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_06BConfig(**config_dict)
@@ -775,7 +814,7 @@ class Qwen3_06B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module):
+class Qwen3_06B_ACE15(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_06B_ACE15_Config(**config_dict)
@@ -784,7 +823,7 @@ class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module):
+class Qwen3_2B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_2B_ACE15_lm_Config(**config_dict)
@@ -793,10 +832,7 @@ class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-    def logits(self, x):
-        return torch.nn.functional.linear(x[:, -1:], self.model.embed_tokens.weight.to(x), None)
-
-class Qwen3_4B(BaseLlama, torch.nn.Module):
+class Qwen3_4B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_4BConfig(**config_dict)
@@ -805,7 +841,16 @@ class Qwen3_4B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_8B(BaseLlama, torch.nn.Module):
+class Qwen3_4B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_4B_ACE15_lm_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+class Qwen3_8B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_8BConfig(**config_dict)

comfy/text_encoders/lt.py

@@ -25,7 +25,7 @@ def ltxv_te(*args, **kwargs):
 class Gemma3_12BTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}

comfy/utils.py

@@ -82,14 +82,12 @@ _TYPES = {
 def load_safetensors(ckpt):
     f = open(ckpt, "rb")
     mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
+    mv = memoryview(mapping)
 
     header_size = struct.unpack("<Q", mapping[:8])[0]
     header = json.loads(mapping[8:8+header_size].decode("utf-8"))
 
-    with warnings.catch_warnings():
-        #We are working with read-only RAM by design
-        warnings.filterwarnings("ignore", message="The given buffer is not writable")
-        data_area = torch.frombuffer(mapping, dtype=torch.uint8)[8 + header_size:]
+    mv = mv[8 + header_size:]
 
     sd = {}
     for name, info in header.items():
@@ -97,7 +95,13 @@ def load_safetensors(ckpt):
             continue
 
         start, end = info["data_offsets"]
-        sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"])
+        if start == end:
+            sd[name] = torch.empty(info["shape"], dtype =_TYPES[info["dtype"]])
+        else:
+            with warnings.catch_warnings():
+                #We are working with read-only RAM by design
+                warnings.filterwarnings("ignore", message="The given buffer is not writable")
+                sd[name] = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
 
     return sd, header.get("__metadata__", {}),
 
@@ -1372,3 +1376,21 @@ def string_to_seed(data):
             else:
                 crc >>= 1
     return crc ^ 0xFFFFFFFF
+
+def deepcopy_list_dict(obj, memo=None):
+    if memo is None:
+        memo = {}
+
+    obj_id = id(obj)
+    if obj_id in memo:
+        return memo[obj_id]
+
+    if isinstance(obj, dict):
+        res = {deepcopy_list_dict(k, memo): deepcopy_list_dict(v, memo) for k, v in obj.items()}
+    elif isinstance(obj, list):
+        res = [deepcopy_list_dict(i, memo) for i in obj]
+    else:
+        res = obj
+
+    memo[obj_id] = res
+    return res

comfy/weight_adapter/bypass.py

@@ -21,6 +21,7 @@ from typing import Optional, Union
 import torch
 import torch.nn as nn
 
+import comfy.model_management
 from .base import WeightAdapterBase, WeightAdapterTrainBase
 from comfy.patcher_extension import PatcherInjection
 
@@ -181,18 +182,21 @@ class BypassForwardHook:
             )
             return  # Already injected
 
-        # Move adapter weights to module's device to avoid CPU-GPU transfer on every forward
-        device = None
+        # Move adapter weights to compute device (GPU)
+        # Use get_torch_device() instead of module.weight.device because
+        # with offloading, module weights may be on CPU while compute happens on GPU
+        device = comfy.model_management.get_torch_device()
+
+        # Get dtype from module weight if available
         dtype = None
         if hasattr(self.module, "weight") and self.module.weight is not None:
-            device = self.module.weight.device
             dtype = self.module.weight.dtype
-        elif hasattr(self.module, "W_q"):  # Quantized layers might use different attr
-            device = self.module.W_q.device
-            dtype = self.module.W_q.dtype
 
-        if device is not None:
-            self._move_adapter_weights_to_device(device, dtype)
+        # Only use dtype if it's a standard float type, not quantized
+        if dtype is not None and dtype not in (torch.float32, torch.float16, torch.bfloat16):
+            dtype = None
+
+        self._move_adapter_weights_to_device(device, dtype)
 
         self.original_forward = self.module.forward
         self.module.forward = self._bypass_forward

comfy_api/latest/__init__.py

@@ -7,7 +7,7 @@ from comfy_api.internal.singleton import ProxiedSingleton
 from comfy_api.internal.async_to_sync import create_sync_class
 from ._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput
 from ._input_impl import VideoFromFile, VideoFromComponents
-from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL
+from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL, File3D
 from . import _io_public as io
 from . import _ui_public as ui
 from comfy_execution.utils import get_executing_context
@@ -105,6 +105,7 @@ class Types:
     VideoComponents = VideoComponents
     MESH = MESH
     VOXEL = VOXEL
+    File3D = File3D
 
 ComfyAPI = ComfyAPI_latest
 

comfy_api/latest/_input/video_types.py

@@ -34,6 +34,21 @@ class VideoInput(ABC):
         """
         pass
 
+    @abstractmethod
+    def as_trimmed(
+        self,
+        start_time: float | None = None,
+        duration: float | None = None,
+        strict_duration: bool = False,
+    ) -> VideoInput | None:
+        """
+        Create a new VideoInput which is trimmed to have the corresponding start_time and duration
+
+        Returns:
+            A new VideoInput, or None if the result would have negative duration
+        """
+        pass
+
     def get_stream_source(self) -> Union[str, io.BytesIO]:
         """
         Get a streamable source for the video. This allows processing without

comfy_api/latest/_input_impl/video_types.py

@@ -6,6 +6,7 @@ from typing import Optional
 from .._input import AudioInput, VideoInput
 import av
 import io
+import itertools
 import json
 import numpy as np
 import math
@@ -29,7 +30,6 @@ def container_to_output_format(container_format: str | None) -> str | None:
     formats = container_format.split(",")
     return formats[0]
 
-
 def get_open_write_kwargs(
     dest: str | io.BytesIO, container_format: str, to_format: str | None
 ) -> dict:
@@ -57,12 +57,14 @@ class VideoFromFile(VideoInput):
     Class representing video input from a file.
     """
 
-    def __init__(self, file: str | io.BytesIO):
+    def __init__(self, file: str | io.BytesIO, *, start_time: float=0, duration: float=0):
         """
         Initialize the VideoFromFile object based off of either a path on disk or a BytesIO object
         containing the file contents.
         """
         self.__file = file
+        self.__start_time = start_time
+        self.__duration = duration
 
     def get_stream_source(self) -> str | io.BytesIO:
         """
@@ -96,6 +98,16 @@ class VideoFromFile(VideoInput):
         Returns:
             Duration in seconds
         """
+        raw_duration = self._get_raw_duration()
+        if self.__start_time < 0:
+            duration_from_start = min(raw_duration, -self.__start_time)
+        else:
+            duration_from_start = raw_duration - self.__start_time
+        if self.__duration:
+            return min(self.__duration, duration_from_start)
+        return duration_from_start
+
+    def _get_raw_duration(self) -> float:
         if isinstance(self.__file, io.BytesIO):
             self.__file.seek(0)
         with av.open(self.__file, mode="r") as container:
@@ -113,9 +125,13 @@ class VideoFromFile(VideoInput):
             if video_stream and video_stream.average_rate:
                 frame_count = 0
                 container.seek(0)
-                for packet in container.demux(video_stream):
-                    for _ in packet.decode():
-                        frame_count += 1
+                frame_iterator = (
+                    container.decode(video_stream)
+                    if video_stream.codec.capabilities & 0x100
+                    else container.demux(video_stream)
+                )
+                for packet in frame_iterator:
+                    frame_count += 1
                 if frame_count > 0:
                     return float(frame_count / video_stream.average_rate)
 
@@ -131,36 +147,54 @@ class VideoFromFile(VideoInput):
 
         with av.open(self.__file, mode="r") as container:
             video_stream = self._get_first_video_stream(container)
-            # 1. Prefer the frames field if available
-            if video_stream.frames and video_stream.frames > 0:
+            # 1. Prefer the frames field if available and usable
+            if (
+                video_stream.frames
+                and video_stream.frames > 0
+                and not self.__start_time
+                and not self.__duration
+            ):
                 return int(video_stream.frames)
 
             # 2. Try to estimate from duration and average_rate using only metadata
-            if container.duration is not None and video_stream.average_rate:
-                duration_seconds = float(container.duration / av.time_base)
-                estimated_frames = int(round(duration_seconds * float(video_stream.average_rate)))
-                if estimated_frames > 0:
-                    return estimated_frames
-
             if (
                 getattr(video_stream, "duration", None) is not None
                 and getattr(video_stream, "time_base", None) is not None
                 and video_stream.average_rate
             ):
-                duration_seconds = float(video_stream.duration * video_stream.time_base)
+                raw_duration = float(video_stream.duration * video_stream.time_base)
+                if self.__start_time < 0:
+                    duration_from_start = min(raw_duration, -self.__start_time)
+                else:
+                    duration_from_start = raw_duration - self.__start_time
+                duration_seconds = min(self.__duration, duration_from_start)
                 estimated_frames = int(round(duration_seconds * float(video_stream.average_rate)))
                 if estimated_frames > 0:
                     return estimated_frames
 
             # 3. Last resort: decode frames and count them (streaming)
-            frame_count = 0
-            container.seek(0)
-            for packet in container.demux(video_stream):
-                for _ in packet.decode():
-                    frame_count += 1
-
-            if frame_count == 0:
-                raise ValueError(f"Could not determine frame count for file '{self.__file}'")
+            if self.__start_time < 0:
+                start_time = max(self._get_raw_duration() + self.__start_time, 0)
+            else:
+                start_time = self.__start_time
+            frame_count = 1
+            start_pts = int(start_time / video_stream.time_base)
+            end_pts = int((start_time + self.__duration) / video_stream.time_base)
+            container.seek(start_pts, stream=video_stream)
+            frame_iterator = (
+                container.decode(video_stream)
+                if video_stream.codec.capabilities & 0x100
+                else container.demux(video_stream)
+            )
+            for frame in frame_iterator:
+                if frame.pts >= start_pts:
+                    break
+            else:
+                raise ValueError(f"Could not determine frame count for file '{self.__file}'\nNo frames exist for start_time {self.__start_time}")
+            for frame in frame_iterator:
+                if frame.pts >= end_pts:
+                    break
+                frame_count += 1
             return frame_count
 
     def get_frame_rate(self) -> Fraction:
@@ -199,9 +233,21 @@ class VideoFromFile(VideoInput):
             return container.format.name
 
     def get_components_internal(self, container: InputContainer) -> VideoComponents:
+        video_stream = self._get_first_video_stream(container)
+        if self.__start_time < 0:
+            start_time = max(self._get_raw_duration() + self.__start_time, 0)
+        else:
+            start_time = self.__start_time
         # Get video frames
         frames = []
-        for frame in container.decode(video=0):
+        start_pts = int(start_time / video_stream.time_base)
+        end_pts = int((start_time + self.__duration) / video_stream.time_base)
+        container.seek(start_pts, stream=video_stream)
+        for frame in container.decode(video_stream):
+            if frame.pts < start_pts:
+                continue
+            if self.__duration and frame.pts >= end_pts:
+                break
             img = frame.to_ndarray(format='rgb24')  # shape: (H, W, 3)
             img = torch.from_numpy(img) / 255.0  # shape: (H, W, 3)
             frames.append(img)
@@ -209,31 +255,44 @@ class VideoFromFile(VideoInput):
         images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0)
 
         # Get frame rate
-        video_stream = next(s for s in container.streams if s.type == 'video')
-        frame_rate = Fraction(video_stream.average_rate) if video_stream and video_stream.average_rate else Fraction(1)
+        frame_rate = Fraction(video_stream.average_rate) if video_stream.average_rate else Fraction(1)
 
         # Get audio if available
         audio = None
-        try:
-            container.seek(0)  # Reset the container to the beginning
-            for stream in container.streams:
-                if stream.type != 'audio':
-                    continue
-                assert isinstance(stream, av.AudioStream)
-                audio_frames = []
-                for packet in container.demux(stream):
-                    for frame in packet.decode():
-                        assert isinstance(frame, av.AudioFrame)
-                        audio_frames.append(frame.to_ndarray())  # shape: (channels, samples)
-                if len(audio_frames) > 0:
-                    audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
-                    audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
-                    audio = AudioInput({
-                        "waveform": audio_tensor,
-                        "sample_rate": int(stream.sample_rate) if stream.sample_rate else 1,
-                    })
-        except StopIteration:
-            pass  # No audio stream
+        container.seek(start_pts, stream=video_stream)
+        # Use last stream for consistency
+        if len(container.streams.audio):
+            audio_stream = container.streams.audio[-1]
+            audio_frames = []
+            resample = av.audio.resampler.AudioResampler(format='fltp').resample
+            frames = itertools.chain.from_iterable(
+                map(resample, container.decode(audio_stream))
+            )
+
+            has_first_frame = False
+            for frame in frames:
+                offset_seconds = start_time - frame.pts * audio_stream.time_base
+                to_skip = int(offset_seconds * audio_stream.sample_rate)
+                if to_skip < frame.samples:
+                    has_first_frame = True
+                    break
+            if has_first_frame:
+                audio_frames.append(frame.to_ndarray()[..., to_skip:])
+
+            for frame in frames:
+                if frame.time > start_time + self.__duration:
+                    break
+                audio_frames.append(frame.to_ndarray())  # shape: (channels, samples)
+            if len(audio_frames) > 0:
+                audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
+                if self.__duration:
+                    audio_data = audio_data[..., :int(self.__duration * audio_stream.sample_rate)]
+
+                audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
+                audio = AudioInput({
+                    "waveform": audio_tensor,
+                    "sample_rate": int(audio_stream.sample_rate) if audio_stream.sample_rate else 1,
+                })
 
         metadata = container.metadata
         return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata)
@@ -250,7 +309,7 @@ class VideoFromFile(VideoInput):
         path: str | io.BytesIO,
         format: VideoContainer = VideoContainer.AUTO,
         codec: VideoCodec = VideoCodec.AUTO,
-        metadata: Optional[dict] = None
+        metadata: Optional[dict] = None,
     ):
         if isinstance(self.__file, io.BytesIO):
             self.__file.seek(0)  # Reset the BytesIO object to the beginning
@@ -262,15 +321,14 @@ class VideoFromFile(VideoInput):
                 reuse_streams = False
             if codec != VideoCodec.AUTO and codec != video_encoding and video_encoding is not None:
                 reuse_streams = False
+            if self.__start_time or self.__duration:
+                reuse_streams = False
 
             if not reuse_streams:
                 components = self.get_components_internal(container)
                 video = VideoFromComponents(components)
                 return video.save_to(
-                    path,
-                    format=format,
-                    codec=codec,
-                    metadata=metadata
+                    path, format=format, codec=codec, metadata=metadata
                 )
 
             streams = container.streams
@@ -304,10 +362,21 @@ class VideoFromFile(VideoInput):
                         output_container.mux(packet)
 
     def _get_first_video_stream(self, container: InputContainer):
-        video_stream = next((s for s in container.streams if s.type == "video"), None)
-        if video_stream is None:
-            raise ValueError(f"No video stream found in file '{self.__file}'")
-        return video_stream
+        if len(container.streams.video):
+            return container.streams.video[0]
+        raise ValueError(f"No video stream found in file '{self.__file}'")
+
+    def as_trimmed(
+        self, start_time: float = 0, duration: float = 0, strict_duration: bool = True
+    ) -> VideoInput | None:
+        trimmed = VideoFromFile(
+            self.get_stream_source(),
+            start_time=start_time + self.__start_time,
+            duration=duration,
+        )
+        if trimmed.get_duration() < duration and strict_duration:
+            return None
+        return trimmed
 
 
 class VideoFromComponents(VideoInput):
@@ -322,7 +391,7 @@ class VideoFromComponents(VideoInput):
         return VideoComponents(
             images=self.__components.images,
             audio=self.__components.audio,
-            frame_rate=self.__components.frame_rate
+            frame_rate=self.__components.frame_rate,
         )
 
     def save_to(
@@ -330,7 +399,7 @@ class VideoFromComponents(VideoInput):
         path: str,
         format: VideoContainer = VideoContainer.AUTO,
         codec: VideoCodec = VideoCodec.AUTO,
-        metadata: Optional[dict] = None
+        metadata: Optional[dict] = None,
     ):
         if format != VideoContainer.AUTO and format != VideoContainer.MP4:
             raise ValueError("Only MP4 format is supported for now")
@@ -357,7 +426,10 @@ class VideoFromComponents(VideoInput):
             audio_stream: Optional[av.AudioStream] = None
             if self.__components.audio:
                 audio_sample_rate = int(self.__components.audio['sample_rate'])
-                audio_stream = output.add_stream('aac', rate=audio_sample_rate)
+                waveform = self.__components.audio['waveform']
+                waveform = waveform[0, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])]
+                layout = {1: 'mono', 2: 'stereo', 6: '5.1'}.get(waveform.shape[0], 'stereo')
+                audio_stream = output.add_stream('aac', rate=audio_sample_rate, layout=layout)
 
             # Encode video
             for i, frame in enumerate(self.__components.images):
@@ -372,12 +444,21 @@ class VideoFromComponents(VideoInput):
             output.mux(packet)
 
             if audio_stream and self.__components.audio:
-                waveform = self.__components.audio['waveform']
-                waveform = waveform[:, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])]
-                frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().cpu().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo')
+                frame = av.AudioFrame.from_ndarray(waveform.float().cpu().numpy(), format='fltp', layout=layout)
                 frame.sample_rate = audio_sample_rate
                 frame.pts = 0
                 output.mux(audio_stream.encode(frame))
 
                 # Flush encoder
                 output.mux(audio_stream.encode(None))
+
+    def as_trimmed(
+        self,
+        start_time: float | None = None,
+        duration: float | None = None,
+        strict_duration: bool = True,
+    ) -> VideoInput | None:
+        if self.get_duration() < start_time + duration:
+            return None
+        #TODO Consider tracking duration and trimming at time of save?
+        return VideoFromFile(self.get_stream_source(), start_time=start_time, duration=duration)

comfy_api/latest/_io.py

@@ -27,7 +27,47 @@ if TYPE_CHECKING:
 from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class,
     prune_dict, shallow_clone_class)
 from comfy_execution.graph_utils import ExecutionBlocker
-from ._util import MESH, VOXEL, SVG as _SVG
+from ._util import MESH, VOXEL, SVG as _SVG, File3D
+
+
+class EmptyInputSentinel:
+    """
+    Sentinel class indicating an empty/missing input.
+
+    Use the class itself (not an instance) as the sentinel.
+    Compare using 'is' or 'is not' only.
+    """
+
+    def __new__(cls):
+        raise TypeError("EmptyInputSentinel cannot be instantiated, use the class itself")
+
+    def __init_subclass__(cls, **kwargs):
+        raise TypeError("EmptyInputSentinel cannot be subclassed")
+
+    @classmethod
+    def __class_getitem__(cls, item):
+        raise TypeError("EmptyInputSentinel cannot be subscripted")
+
+    def __repr__(self):
+        return "<EmptyInput>"
+
+    def __bool__(self):
+        raise TypeError("EmptyInputSentinel cannot be used in boolean context")
+
+    def __eq__(self, other):
+        raise TypeError("EmptyInputSentinel cannot be compared with ==, use 'is' instead")
+
+    def __ne__(self, other):
+        raise TypeError("EmptyInputSentinel cannot be compared with !=, use 'is not' instead")
+
+    def __hash__(self):
+        raise TypeError("EmptyInputSentinel cannot be hashed")
+
+    def __iter__(self):
+        raise TypeError("EmptyInputSentinel cannot be iterated")
+
+    def __len__(self):
+        raise TypeError("EmptyInputSentinel has no length")
 
 
 class FolderType(str, Enum):
@@ -667,6 +707,49 @@ class Voxel(ComfyTypeIO):
 class Mesh(ComfyTypeIO):
     Type = MESH
 
+
+@comfytype(io_type="FILE_3D")
+class File3DAny(ComfyTypeIO):
+    """General 3D file type - accepts any supported 3D format."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_GLB")
+class File3DGLB(ComfyTypeIO):
+    """GLB format 3D file - binary glTF, best for web and cross-platform."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_GLTF")
+class File3DGLTF(ComfyTypeIO):
+    """GLTF format 3D file - JSON-based glTF with external resources."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_FBX")
+class File3DFBX(ComfyTypeIO):
+    """FBX format 3D file - best for game engines and animation."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_OBJ")
+class File3DOBJ(ComfyTypeIO):
+    """OBJ format 3D file - simple geometry format."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_STL")
+class File3DSTL(ComfyTypeIO):
+    """STL format 3D file - best for 3D printing."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_USDZ")
+class File3DUSDZ(ComfyTypeIO):
+    """USDZ format 3D file - Apple AR format."""
+    Type = File3D
+
+
 @comfytype(io_type="HOOKS")
 class Hooks(ComfyTypeIO):
     if TYPE_CHECKING:
@@ -2037,6 +2120,13 @@ __all__ = [
     "LossMap",
     "Voxel",
     "Mesh",
+    "File3DAny",
+    "File3DGLB",
+    "File3DGLTF",
+    "File3DFBX",
+    "File3DOBJ",
+    "File3DSTL",
+    "File3DUSDZ",
     "Hooks",
     "HookKeyframes",
     "TimestepsRange",
@@ -2060,6 +2150,7 @@ __all__ = [
     "DynamicCombo",
     "Autogrow",
     # Other classes
+    "EmptyInputSentinel",
     "HiddenHolder",
     "Hidden",
     "NodeInfoV1",

comfy_api/latest/_util/__init__.py

@@ -1,5 +1,5 @@
 from .video_types import VideoContainer, VideoCodec, VideoComponents
-from .geometry_types import VOXEL, MESH
+from .geometry_types import VOXEL, MESH, File3D
 from .image_types import SVG
 
 __all__ = [
@@ -9,5 +9,6 @@ __all__ = [
     "VideoComponents",
     "VOXEL",
     "MESH",
+    "File3D",
     "SVG",
 ]

comfy_api/latest/_util/geometry_types.py

@@ -1,3 +1,8 @@
+import shutil
+from io import BytesIO
+from pathlib import Path
+from typing import IO
+
 import torch
 
 
@@ -10,3 +15,75 @@ class MESH:
     def __init__(self, vertices: torch.Tensor, faces: torch.Tensor):
         self.vertices = vertices
         self.faces = faces
+
+
+class File3D:
+    """Class representing a 3D file from a file path or binary stream.
+
+    Supports both disk-backed (file path) and memory-backed (BytesIO) storage.
+    """
+
+    def __init__(self, source: str | IO[bytes], file_format: str = ""):
+        self._source = source
+        self._format = file_format or self._infer_format()
+
+    def _infer_format(self) -> str:
+        if isinstance(self._source, str):
+            return Path(self._source).suffix.lstrip(".").lower()
+        return ""
+
+    @property
+    def format(self) -> str:
+        return self._format
+
+    @format.setter
+    def format(self, value: str) -> None:
+        self._format = value.lstrip(".").lower() if value else ""
+
+    @property
+    def is_disk_backed(self) -> bool:
+        return isinstance(self._source, str)
+
+    def get_source(self) -> str | IO[bytes]:
+        if isinstance(self._source, str):
+            return self._source
+        if hasattr(self._source, "seek"):
+            self._source.seek(0)
+        return self._source
+
+    def get_data(self) -> BytesIO:
+        if isinstance(self._source, str):
+            with open(self._source, "rb") as f:
+                result = BytesIO(f.read())
+            return result
+        if hasattr(self._source, "seek"):
+            self._source.seek(0)
+        if isinstance(self._source, BytesIO):
+            return self._source
+        return BytesIO(self._source.read())
+
+    def save_to(self, path: str) -> str:
+        dest = Path(path)
+        dest.parent.mkdir(parents=True, exist_ok=True)
+
+        if isinstance(self._source, str):
+            if Path(self._source).resolve() != dest.resolve():
+                shutil.copy2(self._source, dest)
+        else:
+            if hasattr(self._source, "seek"):
+                self._source.seek(0)
+            with open(dest, "wb") as f:
+                f.write(self._source.read())
+        return str(dest)
+
+    def get_bytes(self) -> bytes:
+        if isinstance(self._source, str):
+            return Path(self._source).read_bytes()
+        if hasattr(self._source, "seek"):
+            self._source.seek(0)
+        return self._source.read()
+
+    def __repr__(self) -> str:
+        if isinstance(self._source, str):
+            return f"File3D(source={self._source!r}, format={self._format!r})"
+        return f"File3D(<stream>, format={self._format!r})"

comfy_api_nodes/apis/__init__.py

@@ -1197,12 +1197,6 @@ class KlingImageGenImageReferenceType(str, Enum):
     face = 'face'
 
 
-class KlingImageGenModelName(str, Enum):
-    kling_v1 = 'kling-v1'
-    kling_v1_5 = 'kling-v1-5'
-    kling_v2 = 'kling-v2'
-
-
 class KlingImageGenerationsRequest(BaseModel):
     aspect_ratio: Optional[KlingImageGenAspectRatio] = '16:9'
     callback_url: Optional[AnyUrl] = Field(
@@ -1218,7 +1212,7 @@ class KlingImageGenerationsRequest(BaseModel):
         0.5, description='Reference intensity for user-uploaded images', ge=0.0, le=1.0
     )
     image_reference: Optional[KlingImageGenImageReferenceType] = None
-    model_name: Optional[KlingImageGenModelName] = 'kling-v1'
+    model_name: str = Field(...)
     n: Optional[int] = Field(1, description='Number of generated images', ge=1, le=9)
     negative_prompt: Optional[str] = Field(
         None, description='Negative text prompt', max_length=200

comfy_api_nodes/apis/kling.py

@@ -1,12 +1,22 @@
 from pydantic import BaseModel, Field
 
 
+class MultiPromptEntry(BaseModel):
+    index: int = Field(...)
+    prompt: str = Field(...)
+    duration: str = Field(...)
+
+
 class OmniProText2VideoRequest(BaseModel):
     model_name: str = Field(..., description="kling-video-o1")
     aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'")
     duration: str = Field(..., description="'5' or '10'")
     prompt: str = Field(...)
     mode: str = Field("pro")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
+    sound: str = Field(..., description="'on' or 'off'")
 
 
 class OmniParamImage(BaseModel):
@@ -26,6 +36,10 @@ class OmniProFirstLastFrameRequest(BaseModel):
     duration: str = Field(..., description="'5' or '10'")
     prompt: str = Field(...)
     mode: str = Field("pro")
+    sound: str | None = Field(None, description="'on' or 'off'")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
 
 
 class OmniProReferences2VideoRequest(BaseModel):
@@ -38,6 +52,10 @@ class OmniProReferences2VideoRequest(BaseModel):
     duration: str | None = Field(..., description="From 3 to 10.")
     prompt: str = Field(...)
     mode: str = Field("pro")
+    sound: str | None = Field(None, description="'on' or 'off'")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
 
 
 class TaskStatusVideoResult(BaseModel):
@@ -54,6 +72,7 @@ class TaskStatusImageResult(BaseModel):
 class TaskStatusResults(BaseModel):
     videos: list[TaskStatusVideoResult] | None = Field(None)
     images: list[TaskStatusImageResult] | None = Field(None)
+    series_images: list[TaskStatusImageResult] | None = Field(None)
 
 
 class TaskStatusResponseData(BaseModel):
@@ -77,31 +96,42 @@ class OmniImageParamImage(BaseModel):
 
 
 class OmniProImageRequest(BaseModel):
-    model_name: str = Field(..., description="kling-image-o1")
-    resolution: str = Field(..., description="'1k' or '2k'")
+    model_name: str = Field(...)
+    resolution: str = Field(...)
     aspect_ratio: str | None = Field(...)
     prompt: str = Field(...)
     mode: str = Field("pro")
     n: int | None = Field(1, le=9)
     image_list: list[OmniImageParamImage] | None = Field(..., max_length=10)
+    result_type: str | None = Field(None, description="Set to 'series' for series generation")
+    series_amount: int | None = Field(None, ge=2, le=9, description="Number of images in a series")
 
 
 class TextToVideoWithAudioRequest(BaseModel):
-    model_name: str = Field(..., description="kling-v2-6")
+    model_name: str = Field(...)
     aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'")
-    duration: str = Field(..., description="'5' or '10'")
-    prompt: str = Field(...)
+    duration: str = Field(...)
+    prompt: str | None = Field(...)
+    negative_prompt: str | None = Field(None)
     mode: str = Field("pro")
     sound: str = Field(..., description="'on' or 'off'")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
 
 
 class ImageToVideoWithAudioRequest(BaseModel):
-    model_name: str = Field(..., description="kling-v2-6")
+    model_name: str = Field(...)
     image: str = Field(...)
-    duration: str = Field(..., description="'5' or '10'")
-    prompt: str = Field(...)
+    image_tail: str | None = Field(None)
+    duration: str = Field(...)
+    prompt: str | None = Field(...)
+    negative_prompt: str | None = Field(None)
     mode: str = Field("pro")
     sound: str = Field(..., description="'on' or 'off'")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
 
 
 class MotionControlRequest(BaseModel):

comfy_api_nodes/apis/meshy.py

@@ -109,14 +109,19 @@ class MeshyTextureRequest(BaseModel):
 
 class MeshyModelsUrls(BaseModel):
     glb: str = Field("")
+    fbx: str = Field("")
+    usdz: str = Field("")
+    obj: str = Field("")
 
 
 class MeshyRiggedModelsUrls(BaseModel):
     rigged_character_glb_url: str = Field("")
+    rigged_character_fbx_url: str = Field("")
 
 
 class MeshyAnimatedModelsUrls(BaseModel):
     animation_glb_url: str = Field("")
+    animation_fbx_url: str = Field("")
 
 
 class MeshyResultTextureUrls(BaseModel):

comfy_api_nodes/nodes_hunyuan3d.py

@@ -1,5 +1,3 @@
-import os
-
 from typing_extensions import override
 
 from comfy_api.latest import IO, ComfyExtension, Input
@@ -14,7 +12,7 @@ from comfy_api_nodes.apis.hunyuan3d import (
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
-    download_url_to_bytesio,
+    download_url_to_file_3d,
     downscale_image_tensor_by_max_side,
     poll_op,
     sync_op,
@@ -22,14 +20,13 @@ from comfy_api_nodes.util import (
     validate_image_dimensions,
     validate_string,
 )
-from folder_paths import get_output_directory
 
 
-def get_glb_obj_from_response(response_objs: list[ResultFile3D]) -> ResultFile3D:
+def get_file_from_response(response_objs: list[ResultFile3D], file_type: str) -> ResultFile3D | None:
     for i in response_objs:
-        if i.Type.lower() == "glb":
+        if i.Type.lower() == file_type.lower():
             return i
-    raise ValueError("No GLB file found in response. Please report this to the developers.")
+    return None
 
 
 class TencentTextToModelNode(IO.ComfyNode):
@@ -74,7 +71,9 @@ class TencentTextToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DOBJ.Output(display_name="OBJ"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -124,19 +123,20 @@ class TencentTextToModelNode(IO.ComfyNode):
         )
         if response.Error:
             raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
+        task_id = response.JobId
         result = await poll_op(
             cls,
             ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro/query", method="POST"),
-            data=To3DProTaskQueryRequest(JobId=response.JobId),
+            data=To3DProTaskQueryRequest(JobId=task_id),
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
-        model_file = f"hunyuan_model_{response.JobId}.glb"
-        await download_url_to_bytesio(
-            get_glb_obj_from_response(result.ResultFile3Ds).Url,
-            os.path.join(get_output_directory(), model_file),
+        glb_result = get_file_from_response(result.ResultFile3Ds, "glb")
+        obj_result = get_file_from_response(result.ResultFile3Ds, "obj")
+        file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None
+        return IO.NodeOutput(
+            file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None
         )
-        return IO.NodeOutput(model_file)
 
 
 class TencentImageToModelNode(IO.ComfyNode):
@@ -184,7 +184,9 @@ class TencentImageToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DOBJ.Output(display_name="OBJ"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -269,19 +271,20 @@ class TencentImageToModelNode(IO.ComfyNode):
         )
         if response.Error:
             raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
+        task_id = response.JobId
         result = await poll_op(
             cls,
             ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro/query", method="POST"),
-            data=To3DProTaskQueryRequest(JobId=response.JobId),
+            data=To3DProTaskQueryRequest(JobId=task_id),
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
-        model_file = f"hunyuan_model_{response.JobId}.glb"
-        await download_url_to_bytesio(
-            get_glb_obj_from_response(result.ResultFile3Ds).Url,
-            os.path.join(get_output_directory(), model_file),
+        glb_result = get_file_from_response(result.ResultFile3Ds, "glb")
+        obj_result = get_file_from_response(result.ResultFile3Ds, "obj")
+        file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None
+        return IO.NodeOutput(
+            file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None
         )
-        return IO.NodeOutput(model_file)
 
 
 class TencentHunyuan3DExtension(ComfyExtension):

comfy_api_nodes/nodes_meshy.py

@@ -1,5 +1,3 @@
-import os
-
 from typing_extensions import override
 
 from comfy_api.latest import IO, ComfyExtension, Input
@@ -20,13 +18,12 @@ from comfy_api_nodes.apis.meshy import (
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
-    download_url_to_bytesio,
+    download_url_to_file_3d,
     poll_op,
     sync_op,
     upload_images_to_comfyapi,
     validate_string,
 )
-from folder_paths import get_output_directory
 
 
 class MeshyTextToModelNode(IO.ComfyNode):
@@ -79,8 +76,10 @@ class MeshyTextToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -122,16 +121,20 @@ class MeshyTextToModelNode(IO.ComfyNode):
                 seed=seed,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyRefineNode(IO.ComfyNode):
@@ -167,8 +170,10 @@ class MeshyRefineNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -210,16 +215,20 @@ class MeshyRefineNode(IO.ComfyNode):
                 ai_model=model,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyImageToModelNode(IO.ComfyNode):
@@ -303,8 +312,10 @@ class MeshyImageToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -368,16 +379,20 @@ class MeshyImageToModelNode(IO.ComfyNode):
                 seed=seed,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/image-to-3d/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/image-to-3d/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyMultiImageToModelNode(IO.ComfyNode):
@@ -464,8 +479,10 @@ class MeshyMultiImageToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -531,16 +548,20 @@ class MeshyMultiImageToModelNode(IO.ComfyNode):
                 seed=seed,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/multi-image-to-3d/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/multi-image-to-3d/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyRigModelNode(IO.ComfyNode):
@@ -571,8 +592,10 @@ class MeshyRigModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_RIGGED_TASK_ID").Output(display_name="rig_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -606,18 +629,20 @@ class MeshyRigModelNode(IO.ComfyNode):
                 texture_image_url=texture_image_url,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/rigging/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/rigging/{task_id}"),
             response_model=MeshyRiggedResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(
-            result.result.rigged_character_glb_url, os.path.join(get_output_directory(), model_file)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.result.rigged_character_glb_url, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.result.rigged_character_fbx_url, "fbx", task_id=task_id),
         )
-        return IO.NodeOutput(model_file, response.result)
 
 
 class MeshyAnimateModelNode(IO.ComfyNode):
@@ -640,7 +665,9 @@ class MeshyAnimateModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -669,16 +696,19 @@ class MeshyAnimateModelNode(IO.ComfyNode):
                 action_id=action_id,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/animations/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/animations/{task_id}"),
             response_model=MeshyAnimationResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.result.animation_glb_url, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            await download_url_to_file_3d(result.result.animation_glb_url, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.result.animation_fbx_url, "fbx", task_id=task_id),
+        )
 
 
 class MeshyTextureNode(IO.ComfyNode):
@@ -715,8 +745,10 @@ class MeshyTextureNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -760,16 +792,20 @@ class MeshyTextureNode(IO.ComfyNode):
                 image_style_url=image_style_url,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/retexture/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/retexture/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyExtension(ComfyExtension):

comfy_api_nodes/nodes_moonvalley.py

@@ -219,8 +219,8 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode):
                 ),
                 IO.Int.Input(
                     "steps",
-                    default=33,
-                    min=1,
+                    default=80,
+                    min=75,  # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0)
                     max=100,
                     step=1,
                     tooltip="Number of denoising steps",
@@ -340,8 +340,8 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode):
                 ),
                 IO.Int.Input(
                     "steps",
-                    default=33,
-                    min=1,
+                    default=60,
+                    min=60,  # steps should be greater or equal to cooldown_steps(36) + warmup_steps(24)
                     max=100,
                     step=1,
                     display_mode=IO.NumberDisplay.number,
@@ -370,7 +370,7 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode):
         video: Input.Video | None = None,
         control_type: str = "Motion Transfer",
         motion_intensity: int | None = 100,
-        steps=33,
+        steps=60,
         prompt_adherence=4.5,
     ) -> IO.NodeOutput:
         validated_video = validate_video_to_video_input(video)
@@ -465,8 +465,8 @@ class MoonvalleyTxt2VideoNode(IO.ComfyNode):
                 ),
                 IO.Int.Input(
                     "steps",
-                    default=33,
-                    min=1,
+                    default=80,
+                    min=75,  # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0)
                     max=100,
                     step=1,
                     tooltip="Inference steps",

comfy_api_nodes/nodes_rodin.py

@@ -10,7 +10,6 @@ import folder_paths as comfy_paths
 import os
 import logging
 import math
-from typing import Optional
 from io import BytesIO
 from typing_extensions import override
 from PIL import Image
@@ -28,8 +27,9 @@ from comfy_api_nodes.util import (
     poll_op,
     ApiEndpoint,
     download_url_to_bytesio,
+    download_url_to_file_3d,
 )
-from comfy_api.latest import ComfyExtension, IO
+from comfy_api.latest import ComfyExtension, IO, Types
 
 
 COMMON_PARAMETERS = [
@@ -177,7 +177,7 @@ def check_rodin_status(response: Rodin3DCheckStatusResponse) -> str:
         return "DONE"
     return "Generating"
 
-def extract_progress(response: Rodin3DCheckStatusResponse) -> Optional[int]:
+def extract_progress(response: Rodin3DCheckStatusResponse) -> int | None:
     if not response.jobs:
         return None
     completed_count = sum(1 for job in response.jobs if job.status == JobStatus.Done)
@@ -207,17 +207,25 @@ async def get_rodin_download_list(uuid: str, cls: type[IO.ComfyNode]) -> Rodin3D
     )
 
 
-async def download_files(url_list, task_uuid: str):
+async def download_files(url_list, task_uuid: str) -> tuple[str | None, Types.File3D | None]:
     result_folder_name = f"Rodin3D_{task_uuid}"
     save_path = os.path.join(comfy_paths.get_output_directory(), result_folder_name)
     os.makedirs(save_path, exist_ok=True)
     model_file_path = None
+    file_3d = None
+
     for i in url_list.list:
         file_path = os.path.join(save_path, i.name)
-        if file_path.endswith(".glb"):
+        if i.name.lower().endswith(".glb"):
             model_file_path = os.path.join(result_folder_name, i.name)
-        await download_url_to_bytesio(i.url, file_path)
-    return model_file_path
+            file_3d = await download_url_to_file_3d(i.url, "glb")
+            # Save to disk for backward compatibility
+            with open(file_path, "wb") as f:
+                f.write(file_3d.get_bytes())
+        else:
+            await download_url_to_bytesio(i.url, file_path)
+
+    return model_file_path, file_3d
 
 
 class Rodin3D_Regular(IO.ComfyNode):
@@ -234,7 +242,10 @@ class Rodin3D_Regular(IO.ComfyNode):
                 IO.Image.Input("Images"),
                 *COMMON_PARAMETERS,
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -271,9 +282,9 @@ class Rodin3D_Regular(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3D_Detail(IO.ComfyNode):
@@ -290,7 +301,10 @@ class Rodin3D_Detail(IO.ComfyNode):
                 IO.Image.Input("Images"),
                 *COMMON_PARAMETERS,
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -327,9 +341,9 @@ class Rodin3D_Detail(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3D_Smooth(IO.ComfyNode):
@@ -346,7 +360,10 @@ class Rodin3D_Smooth(IO.ComfyNode):
                 IO.Image.Input("Images"),
                 *COMMON_PARAMETERS,
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -382,9 +399,9 @@ class Rodin3D_Smooth(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3D_Sketch(IO.ComfyNode):
@@ -408,7 +425,10 @@ class Rodin3D_Sketch(IO.ComfyNode):
                     optional=True,
                 ),
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -441,9 +461,9 @@ class Rodin3D_Sketch(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3D_Gen2(IO.ComfyNode):
@@ -475,7 +495,10 @@ class Rodin3D_Gen2(IO.ComfyNode):
                 ),
                 IO.Boolean.Input("TAPose", default=False),
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -511,9 +534,9 @@ class Rodin3D_Gen2(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3DExtension(ComfyExtension):

comfy_api_nodes/nodes_tripo.py

@@ -1,10 +1,6 @@
-import os
-from typing import Optional
-
-import torch
 from typing_extensions import override
 
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api_nodes.apis.tripo import (
     TripoAnimateRetargetRequest,
     TripoAnimateRigRequest,
@@ -26,12 +22,11 @@ from comfy_api_nodes.apis.tripo import (
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
-    download_url_as_bytesio,
+    download_url_to_file_3d,
     poll_op,
     sync_op,
     upload_images_to_comfyapi,
 )
-from folder_paths import get_output_directory
 
 
 def get_model_url_from_response(response: TripoTaskResponse) -> str:
@@ -45,7 +40,7 @@ def get_model_url_from_response(response: TripoTaskResponse) -> str:
 async def poll_until_finished(
     node_cls: type[IO.ComfyNode],
     response: TripoTaskResponse,
-    average_duration: Optional[int] = None,
+    average_duration: int | None = None,
 ) -> IO.NodeOutput:
     """Polls the Tripo API endpoint until the task reaches a terminal state, then returns the response."""
     if response.code != 0:
@@ -69,12 +64,8 @@ async def poll_until_finished(
     )
     if response_poll.data.status == TripoTaskStatus.SUCCESS:
         url = get_model_url_from_response(response_poll)
-        bytesio = await download_url_as_bytesio(url)
-        # Save the downloaded model file
-        model_file = f"tripo_model_{task_id}.glb"
-        with open(os.path.join(get_output_directory(), model_file), "wb") as f:
-            f.write(bytesio.getvalue())
-        return IO.NodeOutput(model_file, task_id)
+        file_glb = await download_url_to_file_3d(url, "glb", task_id=task_id)
+        return IO.NodeOutput(f"{task_id}.glb", task_id, file_glb)
     raise RuntimeError(f"Failed to generate mesh: {response_poll}")
 
 
@@ -107,8 +98,9 @@ class TripoTextToModelNode(IO.ComfyNode):
                 IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -155,18 +147,18 @@ class TripoTextToModelNode(IO.ComfyNode):
     async def execute(
         cls,
         prompt: str,
-        negative_prompt: Optional[str] = None,
+        negative_prompt: str | None = None,
         model_version=None,
-        style: Optional[str] = None,
-        texture: Optional[bool] = None,
-        pbr: Optional[bool] = None,
-        image_seed: Optional[int] = None,
-        model_seed: Optional[int] = None,
-        texture_seed: Optional[int] = None,
-        texture_quality: Optional[str] = None,
-        geometry_quality: Optional[str] = None,
-        face_limit: Optional[int] = None,
-        quad: Optional[bool] = None,
+        style: str | None = None,
+        texture: bool | None = None,
+        pbr: bool | None = None,
+        image_seed: int | None = None,
+        model_seed: int | None = None,
+        texture_seed: int | None = None,
+        texture_quality: str | None = None,
+        geometry_quality: str | None = None,
+        face_limit: int | None = None,
+        quad: bool | None = None,
     ) -> IO.NodeOutput:
         style_enum = None if style == "None" else style
         if not prompt:
@@ -232,8 +224,9 @@ class TripoImageToModelNode(IO.ComfyNode):
                 IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -279,19 +272,19 @@ class TripoImageToModelNode(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        image: torch.Tensor,
-        model_version: Optional[str] = None,
-        style: Optional[str] = None,
-        texture: Optional[bool] = None,
-        pbr: Optional[bool] = None,
-        model_seed: Optional[int] = None,
+        image: Input.Image,
+        model_version: str | None = None,
+        style: str | None = None,
+        texture: bool | None = None,
+        pbr: bool | None = None,
+        model_seed: int | None = None,
         orientation=None,
-        texture_seed: Optional[int] = None,
-        texture_quality: Optional[str] = None,
-        geometry_quality: Optional[str] = None,
-        texture_alignment: Optional[str] = None,
-        face_limit: Optional[int] = None,
-        quad: Optional[bool] = None,
+        texture_seed: int | None = None,
+        texture_quality: str | None = None,
+        geometry_quality: str | None = None,
+        texture_alignment: str | None = None,
+        face_limit: int | None = None,
+        quad: bool | None = None,
     ) -> IO.NodeOutput:
         style_enum = None if style == "None" else style
         if image is None:
@@ -368,8 +361,9 @@ class TripoMultiviewToModelNode(IO.ComfyNode):
                 IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -411,21 +405,21 @@ class TripoMultiviewToModelNode(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        image: torch.Tensor,
-        image_left: Optional[torch.Tensor] = None,
-        image_back: Optional[torch.Tensor] = None,
-        image_right: Optional[torch.Tensor] = None,
-        model_version: Optional[str] = None,
-        orientation: Optional[str] = None,
-        texture: Optional[bool] = None,
-        pbr: Optional[bool] = None,
-        model_seed: Optional[int] = None,
-        texture_seed: Optional[int] = None,
-        texture_quality: Optional[str] = None,
-        geometry_quality: Optional[str] = None,
-        texture_alignment: Optional[str] = None,
-        face_limit: Optional[int] = None,
-        quad: Optional[bool] = None,
+        image: Input.Image,
+        image_left: Input.Image | None = None,
+        image_back: Input.Image | None = None,
+        image_right: Input.Image | None = None,
+        model_version: str | None = None,
+        orientation: str | None = None,
+        texture: bool | None = None,
+        pbr: bool | None = None,
+        model_seed: int | None = None,
+        texture_seed: int | None = None,
+        texture_quality: str | None = None,
+        geometry_quality: str | None = None,
+        texture_alignment: str | None = None,
+        face_limit: int | None = None,
+        quad: bool | None = None,
     ) -> IO.NodeOutput:
         if image is None:
             raise RuntimeError("front image for multiview is required")
@@ -487,8 +481,9 @@ class TripoTextureNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -512,11 +507,11 @@ class TripoTextureNode(IO.ComfyNode):
     async def execute(
         cls,
         model_task_id,
-        texture: Optional[bool] = None,
-        pbr: Optional[bool] = None,
-        texture_seed: Optional[int] = None,
-        texture_quality: Optional[str] = None,
-        texture_alignment: Optional[str] = None,
+        texture: bool | None = None,
+        pbr: bool | None = None,
+        texture_seed: int | None = None,
+        texture_quality: str | None = None,
+        texture_alignment: str | None = None,
     ) -> IO.NodeOutput:
         response = await sync_op(
             cls,
@@ -547,8 +542,9 @@ class TripoRefineNode(IO.ComfyNode):
                 IO.Custom("MODEL_TASK_ID").Input("model_task_id", tooltip="Must be a v1.4 Tripo model"),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -583,8 +579,9 @@ class TripoRigNode(IO.ComfyNode):
             category="api node/3d/Tripo",
             inputs=[IO.Custom("MODEL_TASK_ID").Input("original_model_task_id")],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("RIG_TASK_ID").Output(display_name="rig task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -642,8 +639,9 @@ class TripoRetargetNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("RETARGET_TASK_ID").Output(display_name="retarget task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,

comfy_api_nodes/util/__init__.py

@@ -28,6 +28,7 @@ from .conversions import (
 from .download_helpers import (
     download_url_as_bytesio,
     download_url_to_bytesio,
+    download_url_to_file_3d,
     download_url_to_image_tensor,
     download_url_to_video_output,
 )
@@ -69,6 +70,7 @@ __all__ = [
     # Download helpers
     "download_url_as_bytesio",
     "download_url_to_bytesio",
+    "download_url_to_file_3d",
     "download_url_to_image_tensor",
     "download_url_to_video_output",
     # Conversions

comfy_api_nodes/util/download_helpers.py

@@ -11,7 +11,8 @@ import torch
 from aiohttp.client_exceptions import ClientError, ContentTypeError
 
 from comfy_api.latest import IO as COMFY_IO
-from comfy_api.latest import InputImpl
+from comfy_api.latest import InputImpl, Types
+from folder_paths import get_output_directory
 
 from . import request_logger
 from ._helpers import (
@@ -261,3 +262,38 @@ def _generate_operation_id(method: str, url: str, attempt: int) -> str:
     except Exception:
         slug = "download"
     return f"{method}_{slug}_try{attempt}_{uuid.uuid4().hex[:8]}"
+
+
+async def download_url_to_file_3d(
+    url: str,
+    file_format: str,
+    *,
+    task_id: str | None = None,
+    timeout: float | None = None,
+    max_retries: int = 5,
+    cls: type[COMFY_IO.ComfyNode] = None,
+) -> Types.File3D:
+    """Downloads a 3D model file from a URL into memory as BytesIO.
+
+    If task_id is provided, also writes the file to disk in the output directory
+    for backward compatibility with the old save-to-disk behavior.
+    """
+    file_format = file_format.lstrip(".").lower()
+    data = BytesIO()
+    await download_url_to_bytesio(
+        url,
+        data,
+        timeout=timeout,
+        max_retries=max_retries,
+        cls=cls,
+    )
+
+    if task_id is not None:
+        # This is only for backward compatability with current behavior when every 3D node is output node
+        # All new API nodes should not use "task_id" and instead users should use "SaveGLB" node to save results
+        output_dir = Path(get_output_directory())
+        output_path = output_dir / f"{task_id}.{file_format}"
+        output_path.write_bytes(data.getvalue())
+        data.seek(0)
+
+    return Types.File3D(source=data, file_format=file_format)

comfy_execution/jobs.py

@@ -20,10 +20,60 @@ class JobStatus:
 
 
 # Media types that can be previewed in the frontend
-PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio'})
+PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio', '3d'})
 
 # 3D file extensions for preview fallback (no dedicated media_type exists)
-THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb'})
+THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb', '.usdz'})
+
+
+def has_3d_extension(filename: str) -> bool:
+    lower = filename.lower()
+    return any(lower.endswith(ext) for ext in THREE_D_EXTENSIONS)
+
+
+def normalize_output_item(item):
+    """Normalize a single output list item for the jobs API.
+
+    Returns the normalized item, or None to exclude it.
+    String items with 3D extensions become {filename, type, subfolder} dicts.
+    """
+    if item is None:
+        return None
+    if isinstance(item, str):
+        if has_3d_extension(item):
+            return {'filename': item, 'type': 'output', 'subfolder': '', 'mediaType': '3d'}
+        return None
+    if isinstance(item, dict):
+        return item
+    return None
+
+
+def normalize_outputs(outputs: dict) -> dict:
+    """Normalize raw node outputs for the jobs API.
+
+    Transforms string 3D filenames into file output dicts and removes
+    None items. All other items (non-3D strings, dicts, etc.) are
+    preserved as-is.
+    """
+    normalized = {}
+    for node_id, node_outputs in outputs.items():
+        if not isinstance(node_outputs, dict):
+            normalized[node_id] = node_outputs
+            continue
+        normalized_node = {}
+        for media_type, items in node_outputs.items():
+            if media_type == 'animated' or not isinstance(items, list):
+                normalized_node[media_type] = items
+                continue
+            normalized_items = []
+            for item in items:
+                if item is None:
+                    continue
+                norm = normalize_output_item(item)
+                normalized_items.append(norm if norm is not None else item)
+            normalized_node[media_type] = normalized_items
+        normalized[node_id] = normalized_node
+    return normalized
 
 
 def _extract_job_metadata(extra_data: dict) -> tuple[Optional[int], Optional[str]]:
@@ -45,9 +95,9 @@ def is_previewable(media_type: str, item: dict) -> bool:
     Maintains backwards compatibility with existing logic.
 
     Priority:
-    1. media_type is 'images', 'video', or 'audio'
+    1. media_type is 'images', 'video', 'audio', or '3d'
     2. format field starts with 'video/' or 'audio/'
-    3. filename has a 3D extension (.obj, .fbx, .gltf, .glb)
+    3. filename has a 3D extension (.obj, .fbx, .gltf, .glb, .usdz)
     """
     if media_type in PREVIEWABLE_MEDIA_TYPES:
         return True
@@ -139,7 +189,7 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs:
     })
 
     if include_outputs:
-        job['outputs'] = outputs
+        job['outputs'] = normalize_outputs(outputs)
         job['execution_status'] = status_info
         job['workflow'] = {
             'prompt': prompt,
@@ -171,18 +221,23 @@ def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]:
                 continue
 
             for item in items:
-                count += 1
-
-                if not isinstance(item, dict):
+                normalized = normalize_output_item(item)
+                if normalized is None:
                     continue
 
-                if preview_output is None and is_previewable(media_type, item):
+                count += 1
+
+                if preview_output is not None:
+                    continue
+
+                if isinstance(normalized, dict) and is_previewable(media_type, normalized):
                     enriched = {
-                        **item,
+                        **normalized,
                         'nodeId': node_id,
-                        'mediaType': media_type
                     }
-                    if item.get('type') == 'output':
+                    if 'mediaType' not in normalized:
+                        enriched['mediaType'] = media_type
+                    if normalized.get('type') == 'output':
                         preview_output = enriched
                     elif fallback_preview is None:
                         fallback_preview = enriched

comfy_extras/nodes_ace.py

@@ -44,13 +44,18 @@ class TextEncodeAceStepAudio15(io.ComfyNode):
                 io.Combo.Input("timesignature", options=['2', '3', '4', '6']),
                 io.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]),
                 io.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]),
+                io.Boolean.Input("generate_audio_codes", default=True, tooltip="Enable the LLM that generates audio codes. This can be slow but will increase the quality of the generated audio. Turn this off if you are giving the model an audio reference.", advanced=True),
+                io.Float.Input("cfg_scale", default=2.0, min=0.0, max=100.0, step=0.1, advanced=True),
+                io.Float.Input("temperature", default=0.85, min=0.0, max=2.0, step=0.01, advanced=True),
+                io.Float.Input("top_p", default=0.9, min=0.0, max=2000.0, step=0.01, advanced=True),
+                io.Int.Input("top_k", default=0, min=0, max=100, advanced=True),
             ],
             outputs=[io.Conditioning.Output()],
         )
 
     @classmethod
-    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale) -> io.NodeOutput:
-        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed)
+    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale, generate_audio_codes, cfg_scale, temperature, top_p, top_k) -> io.NodeOutput:
+        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed, generate_audio_codes=generate_audio_codes, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k)
         conditioning = clip.encode_from_tokens_scheduled(tokens)
         return io.NodeOutput(conditioning)
 
@@ -100,14 +105,15 @@ class EmptyAceStep15LatentAudio(io.ComfyNode):
         latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device())
         return io.NodeOutput({"samples": latent, "type": "audio"})
 
-class ReferenceTimbreAudio(io.ComfyNode):
+class ReferenceAudio(io.ComfyNode):
     @classmethod
     def define_schema(cls):
         return io.Schema(
             node_id="ReferenceTimbreAudio",
+            display_name="Reference Audio",
             category="advanced/conditioning/audio",
             is_experimental=True,
-            description="This node sets the reference audio for timbre (for ace step 1.5)",
+            description="This node sets the reference audio for ace step 1.5",
             inputs=[
                 io.Conditioning.Input("conditioning"),
                 io.Latent.Input("latent", optional=True),
@@ -131,7 +137,7 @@ class AceExtension(ComfyExtension):
             EmptyAceStepLatentAudio,
             TextEncodeAceStepAudio15,
             EmptyAceStep15LatentAudio,
-            ReferenceTimbreAudio,
+            ReferenceAudio,
         ]
 
 async def comfy_entrypoint() -> AceExtension:

comfy_extras/nodes_audio.py

@@ -94,6 +94,19 @@ class VAEEncodeAudio(IO.ComfyNode):
     encode = execute  # TODO: remove
 
 
+def vae_decode_audio(vae, samples, tile=None, overlap=None):
+    if tile is not None:
+        audio = vae.decode_tiled(samples["samples"], tile_y=tile, overlap=overlap).movedim(-1, 1)
+    else:
+        audio = vae.decode(samples["samples"]).movedim(-1, 1)
+
+    std = torch.std(audio, dim=[1, 2], keepdim=True) * 5.0
+    std[std < 1.0] = 1.0
+    audio /= std
+    vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+    return {"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]}
+
+
 class VAEDecodeAudio(IO.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -111,16 +124,33 @@ class VAEDecodeAudio(IO.ComfyNode):
 
     @classmethod
     def execute(cls, vae, samples) -> IO.NodeOutput:
-        audio = vae.decode(samples["samples"]).movedim(-1, 1)
-        std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0
-        std[std < 1.0] = 1.0
-        audio /= std
-        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
-        return IO.NodeOutput({"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]})
+        return IO.NodeOutput(vae_decode_audio(vae, samples))
 
     decode = execute  # TODO: remove
 
 
+class VAEDecodeAudioTiled(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="VAEDecodeAudioTiled",
+            search_aliases=["latent to audio"],
+            display_name="VAE Decode Audio (Tiled)",
+            category="latent/audio",
+            inputs=[
+                IO.Latent.Input("samples"),
+                IO.Vae.Input("vae"),
+                IO.Int.Input("tile_size", default=512, min=32, max=8192, step=8),
+                IO.Int.Input("overlap", default=64, min=0, max=1024, step=8),
+            ],
+            outputs=[IO.Audio.Output()],
+        )
+
+    @classmethod
+    def execute(cls, vae, samples, tile_size, overlap) -> IO.NodeOutput:
+        return IO.NodeOutput(vae_decode_audio(vae, samples, tile_size, overlap))
+
+
 class SaveAudio(IO.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -675,6 +705,7 @@ class AudioExtension(ComfyExtension):
             EmptyLatentAudio,
             VAEEncodeAudio,
             VAEDecodeAudio,
+            VAEDecodeAudioTiled,
             SaveAudio,
             SaveAudioMP3,
             SaveAudioOpus,

comfy_extras/nodes_custom_sampler.py

@@ -622,6 +622,7 @@ class SamplerSASolver(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerSASolver",
+            search_aliases=["sde"],
             category="sampling/custom_sampling/samplers",
             inputs=[
                 io.Model.Input("model"),
@@ -666,6 +667,7 @@ class SamplerSEEDS2(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerSEEDS2",
+            search_aliases=["sde", "exp heun"],
             category="sampling/custom_sampling/samplers",
             inputs=[
                 io.Combo.Input("solver_type", options=["phi_1", "phi_2"]),

comfy_extras/nodes_easycache.py

@@ -9,6 +9,14 @@ if TYPE_CHECKING:
     from uuid import UUID
 
 
+def _extract_tensor(data, output_channels):
+    """Extract tensor from data, handling both single tensors and lists."""
+    if isinstance(data, list):
+        # LTX2 AV tensors: [video, audio]
+        return data[0][:, :output_channels], data[1][:, :output_channels]
+    return data[:, :output_channels], None
+
+
 def easycache_forward_wrapper(executor, *args, **kwargs):
     # get values from args
     transformer_options: dict[str] = args[-1]
@@ -17,7 +25,7 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
         if not transformer_options:
             transformer_options = args[-2]
     easycache: EasyCacheHolder = transformer_options["easycache"]
-    x: torch.Tensor = args[0][:, :easycache.output_channels]
+    x, ax = _extract_tensor(args[0], easycache.output_channels)
     sigmas = transformer_options["sigmas"]
     uuids = transformer_options["uuids"]
     if sigmas is not None and easycache.is_past_end_timestep(sigmas):
@@ -35,7 +43,11 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
         if easycache.skip_current_step and can_apply_cache_diff:
             if easycache.verbose:
                 logging.info(f"EasyCache [verbose] - was marked to skip this step by {easycache.first_cond_uuid}. Present uuids: {uuids}")
-            return easycache.apply_cache_diff(x, uuids)
+            result = easycache.apply_cache_diff(x, uuids)
+            if ax is not None:
+                result_audio = easycache.apply_cache_diff(ax, uuids, is_audio=True)
+                return [result, result_audio]
+            return result
         if easycache.initial_step:
             easycache.first_cond_uuid = uuids[0]
             has_first_cond_uuid = easycache.has_first_cond_uuid(uuids)
@@ -51,13 +63,18 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
                         logging.info(f"EasyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
                     # other conds should also skip this step, and instead use their cached values
                     easycache.skip_current_step = True
-                    return easycache.apply_cache_diff(x, uuids)
+                    result = easycache.apply_cache_diff(x, uuids)
+                    if ax is not None:
+                        result_audio = easycache.apply_cache_diff(ax, uuids, is_audio=True)
+                        return [result, result_audio]
+                    return result
                 else:
                     if easycache.verbose:
                         logging.info(f"EasyCache [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
                     easycache.cumulative_change_rate = 0.0
 
-    output: torch.Tensor = executor(*args, **kwargs)
+    full_output: torch.Tensor = executor(*args, **kwargs)
+    output, audio_output = _extract_tensor(full_output, easycache.output_channels)
     if has_first_cond_uuid and easycache.has_output_prev_norm():
         output_change = (easycache.subsample(output, uuids, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean()
         if easycache.verbose:
@@ -74,13 +91,15 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
             logging.info(f"EasyCache [verbose] - output_change_rate: {output_change_rate}")
     # TODO: allow cache_diff to be offloaded
     easycache.update_cache_diff(output, next_x_prev, uuids)
+    if audio_output is not None:
+        easycache.update_cache_diff(audio_output, ax, uuids, is_audio=True)
     if has_first_cond_uuid:
         easycache.x_prev_subsampled = easycache.subsample(next_x_prev, uuids)
         easycache.output_prev_subsampled = easycache.subsample(output, uuids)
         easycache.output_prev_norm = output.flatten().abs().mean()
         if easycache.verbose:
             logging.info(f"EasyCache [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}")
-    return output
+    return full_output
 
 def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
     # get values from args
@@ -89,8 +108,8 @@ def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
     easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"]
     if easycache.is_past_end_timestep(timestep):
         return executor(*args, **kwargs)
-    # prepare next x_prev
     x: torch.Tensor = args[0][:, :easycache.output_channels]
+    # prepare next x_prev
     next_x_prev = x
     input_change = None
     do_easycache = easycache.should_do_easycache(timestep)
@@ -197,6 +216,7 @@ class EasyCacheHolder:
         self.output_prev_subsampled: torch.Tensor = None
         self.output_prev_norm: torch.Tensor = None
         self.uuid_cache_diffs: dict[UUID, torch.Tensor] = {}
+        self.uuid_cache_diffs_audio: dict[UUID, torch.Tensor] = {}
         self.output_change_rates = []
         self.approx_output_change_rates = []
         self.total_steps_skipped = 0
@@ -245,20 +265,21 @@ class EasyCacheHolder:
     def can_apply_cache_diff(self, uuids: list[UUID]) -> bool:
         return all(uuid in self.uuid_cache_diffs for uuid in uuids)
 
-    def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID]):
-        if self.first_cond_uuid in uuids:
+    def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID], is_audio: bool = False):
+        if self.first_cond_uuid in uuids and not is_audio:
             self.total_steps_skipped += 1
+        cache_diffs = self.uuid_cache_diffs_audio if is_audio else self.uuid_cache_diffs
         batch_offset = x.shape[0] // len(uuids)
         for i, uuid in enumerate(uuids):
             # slice out only what is relevant to this cond
             batch_slice = [slice(i*batch_offset,(i+1)*batch_offset)]
             # if cached dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
-            if x.shape[1:] != self.uuid_cache_diffs[uuid].shape[1:]:
+            if x.shape[1:] != cache_diffs[uuid].shape[1:]:
                 if not self.allow_mismatch:
                     raise ValueError(f"Cached dims {self.uuid_cache_diffs[uuid].shape} don't match x dims {x.shape} - this is no good")
                 slicing = []
                 skip_this_dim = True
-                for dim_u, dim_x in zip(self.uuid_cache_diffs[uuid].shape, x.shape):
+                for dim_u, dim_x in zip(cache_diffs[uuid].shape, x.shape):
                     if skip_this_dim:
                         skip_this_dim = False
                         continue
@@ -270,10 +291,11 @@ class EasyCacheHolder:
                     else:
                         slicing.append(slice(None))
                 batch_slice = batch_slice + slicing
-            x[tuple(batch_slice)] += self.uuid_cache_diffs[uuid].to(x.device)
+            x[tuple(batch_slice)] += cache_diffs[uuid].to(x.device)
         return x
 
-    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]):
+    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID], is_audio: bool = False):
+        cache_diffs = self.uuid_cache_diffs_audio if is_audio else self.uuid_cache_diffs
         # if output dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
         if output.shape[1:] != x.shape[1:]:
             if not self.allow_mismatch:
@@ -293,7 +315,7 @@ class EasyCacheHolder:
         diff = output - x
         batch_offset = diff.shape[0] // len(uuids)
         for i, uuid in enumerate(uuids):
-            self.uuid_cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...]
+            cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...]
 
     def has_first_cond_uuid(self, uuids: list[UUID]) -> bool:
         return self.first_cond_uuid in uuids
@@ -324,6 +346,8 @@ class EasyCacheHolder:
         self.output_prev_norm = None
         del self.uuid_cache_diffs
         self.uuid_cache_diffs = {}
+        del self.uuid_cache_diffs_audio
+        self.uuid_cache_diffs_audio = {}
         self.total_steps_skipped = 0
         self.state_metadata = None
         return self

comfy_extras/nodes_hunyuan3d.py

@@ -618,18 +618,31 @@ class SaveGLB(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="SaveGLB",
+            display_name="Save 3D Model",
             search_aliases=["export 3d model", "save mesh"],
             category="3d",
             is_output_node=True,
             inputs=[
-                IO.Mesh.Input("mesh"),
+                IO.MultiType.Input(
+                    IO.Mesh.Input("mesh"),
+                    types=[
+                        IO.File3DGLB,
+                        IO.File3DGLTF,
+                        IO.File3DOBJ,
+                        IO.File3DFBX,
+                        IO.File3DSTL,
+                        IO.File3DUSDZ,
+                        IO.File3DAny,
+                    ],
+                    tooltip="Mesh or 3D file to save",
+                ),
                 IO.String.Input("filename_prefix", default="mesh/ComfyUI"),
             ],
             hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo]
         )
 
     @classmethod
-    def execute(cls, mesh, filename_prefix) -> IO.NodeOutput:
+    def execute(cls, mesh: Types.MESH | Types.File3D, filename_prefix: str) -> IO.NodeOutput:
         full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory())
         results = []
 
@@ -641,15 +654,27 @@ class SaveGLB(IO.ComfyNode):
                 for x in cls.hidden.extra_pnginfo:
                     metadata[x] = json.dumps(cls.hidden.extra_pnginfo[x])
 
-        for i in range(mesh.vertices.shape[0]):
-            f = f"{filename}_{counter:05}_.glb"
-            save_glb(mesh.vertices[i], mesh.faces[i], os.path.join(full_output_folder, f), metadata)
+        if isinstance(mesh, Types.File3D):
+            # Handle File3D input - save BytesIO data to output folder
+            ext = mesh.format or "glb"
+            f = f"{filename}_{counter:05}_.{ext}"
+            mesh.save_to(os.path.join(full_output_folder, f))
             results.append({
                 "filename": f,
                 "subfolder": subfolder,
                 "type": "output"
             })
-            counter += 1
+        else:
+            # Handle Mesh input - save vertices and faces as GLB
+            for i in range(mesh.vertices.shape[0]):
+                f = f"{filename}_{counter:05}_.glb"
+                save_glb(mesh.vertices[i], mesh.faces[i], os.path.join(full_output_folder, f), metadata)
+                results.append({
+                    "filename": f,
+                    "subfolder": subfolder,
+                    "type": "output"
+                })
+                counter += 1
         return IO.NodeOutput(ui={"3d": results})
 
 

comfy_extras/nodes_latent.py

@@ -391,8 +391,9 @@ class LatentOperationTonemapReinhard(io.ComfyNode):
             latent_vector_magnitude = (torch.linalg.vector_norm(latent, dim=(1)) + 0.0000000001)[:,None]
             normalized_latent = latent / latent_vector_magnitude
 
-            mean = torch.mean(latent_vector_magnitude, dim=(1,2,3), keepdim=True)
-            std = torch.std(latent_vector_magnitude, dim=(1,2,3), keepdim=True)
+            dims = list(range(1, latent_vector_magnitude.ndim))
+            mean = torch.mean(latent_vector_magnitude, dim=dims, keepdim=True)
+            std = torch.std(latent_vector_magnitude, dim=dims, keepdim=True)
 
             top = (std * 5 + mean) * multiplier
 

comfy_extras/nodes_load_3d.py

@@ -1,9 +1,10 @@
 import nodes
 import folder_paths
 import os
+import uuid
 
 from typing_extensions import override
-from comfy_api.latest import IO, ComfyExtension, InputImpl, UI
+from comfy_api.latest import IO, UI, ComfyExtension, InputImpl, Types
 
 from pathlib import Path
 
@@ -44,6 +45,7 @@ class Load3D(IO.ComfyNode):
                 IO.Image.Output(display_name="normal"),
                 IO.Load3DCamera.Output(display_name="camera_info"),
                 IO.Video.Output(display_name="recording_video"),
+                IO.File3DAny.Output(display_name="model_3d"),
             ],
         )
 
@@ -65,7 +67,8 @@ class Load3D(IO.ComfyNode):
 
             video = InputImpl.VideoFromFile(recording_video_path)
 
-        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video)
+        file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file))
+        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video, file_3d)
 
     process = execute  # TODO: remove
 
@@ -81,7 +84,19 @@ class Preview3D(IO.ComfyNode):
             is_experimental=True,
             is_output_node=True,
             inputs=[
-                IO.String.Input("model_file", default="", multiline=False),
+                IO.MultiType.Input(
+                    IO.String.Input("model_file", default="", multiline=False),
+                    types=[
+                        IO.File3DGLB,
+                        IO.File3DGLTF,
+                        IO.File3DFBX,
+                        IO.File3DOBJ,
+                        IO.File3DSTL,
+                        IO.File3DUSDZ,
+                        IO.File3DAny,
+                    ],
+                    tooltip="3D model file or path string",
+                ),
                 IO.Load3DCamera.Input("camera_info", optional=True),
                 IO.Image.Input("bg_image", optional=True),
             ],
@@ -89,10 +104,15 @@ class Preview3D(IO.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, model_file, **kwargs) -> IO.NodeOutput:
+    def execute(cls, model_file: str | Types.File3D, **kwargs) -> IO.NodeOutput:
+        if isinstance(model_file, Types.File3D):
+            filename = f"preview3d_{uuid.uuid4().hex}.{model_file.format}"
+            model_file.save_to(os.path.join(folder_paths.get_output_directory(), filename))
+        else:
+            filename = model_file
         camera_info = kwargs.get("camera_info", None)
         bg_image = kwargs.get("bg_image", None)
-        return IO.NodeOutput(ui=UI.PreviewUI3D(model_file, camera_info, bg_image=bg_image))
+        return IO.NodeOutput(ui=UI.PreviewUI3D(filename, camera_info, bg_image=bg_image))
 
     process = execute  # TODO: remove
 

comfy_extras/nodes_logic.py

@@ -91,6 +91,41 @@ class SoftSwitchNode(io.ComfyNode):
         return io.NodeOutput(on_true if switch else on_false)
 
 
+class OptionalSwitchNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        template = io.MatchType.Template("switch")
+        return io.Schema(
+            node_id="ComfyOptionalSwitchNode",
+            display_name="Optional Switch",
+            category="logic",
+            is_experimental=True,
+            inputs=[
+                io.Boolean.Input("switch"),
+                io.MatchType.Input("on_false", template=template, lazy=True, optional=True),
+                io.MatchType.Input("on_true", template=template, lazy=True, optional=True),
+            ],
+            outputs=[
+                io.MatchType.Output(template=template, display_name="output"),
+            ],
+        )
+
+    @classmethod
+    def check_lazy_status(cls, switch, on_false=MISSING, on_true=MISSING):
+        # Only evaluate the input that corresponds to the switch value
+        if switch and on_true is None:
+            return ["on_true"]
+        if not switch and on_false is None:
+            return ["on_false"]
+
+    @classmethod
+    def execute(cls, switch, on_true=MISSING, on_false=MISSING) -> io.NodeOutput:
+        selected = on_true if switch else on_false
+        if selected is MISSING:
+            return io.NodeOutput(io.EmptyInputSentinel)
+        return io.NodeOutput(selected)
+
+
 class CustomComboNode(io.ComfyNode):
     """
     Frontend node that allows user to write their own options for a combo.
@@ -260,6 +295,7 @@ class LogicExtension(ComfyExtension):
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
         return [
             SwitchNode,
+            OptionalSwitchNode,
             CustomComboNode,
             # SoftSwitchNode,
             # ConvertStringToComboNode,

comfy_extras/nodes_toolkit.py

@@ -0,0 +1,47 @@
+from __future__ import annotations
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, io
+
+
+class CreateList(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        template_matchtype = io.MatchType.Template("type")
+        template_autogrow = io.Autogrow.TemplatePrefix(
+            input=io.MatchType.Input("input", template=template_matchtype),
+            prefix="input",
+        )
+        return io.Schema(
+            node_id="CreateList",
+            display_name="Create List",
+            category="logic",
+            is_input_list=True,
+            search_aliases=["Image Iterator", "Text Iterator", "Iterator"],
+            inputs=[io.Autogrow.Input("inputs", template=template_autogrow)],
+            outputs=[
+                io.MatchType.Output(
+                    template=template_matchtype,
+                    is_output_list=True,
+                    display_name="list",
+                ),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, inputs: io.Autogrow.Type) -> io.NodeOutput:
+        output_list = []
+        for input in inputs.values():
+            output_list += input
+        return io.NodeOutput(output_list)
+
+
+class ToolkitExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            CreateList,
+        ]
+
+
+async def comfy_entrypoint() -> ToolkitExtension:
+    return ToolkitExtension()

comfy_extras/nodes_train.py

@@ -4,6 +4,7 @@ import os
 import numpy as np
 import safetensors
 import torch
+import torch.nn as nn
 import torch.utils.checkpoint
 from tqdm.auto import trange
 from PIL import Image, ImageDraw, ImageFont
@@ -27,6 +28,11 @@ class TrainGuider(comfy_extras.nodes_custom_sampler.Guider_Basic):
     """
     CFGGuider with modifications for training specific logic
     """
+
+    def __init__(self, *args, offloading=False, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.offloading = offloading
+
     def outer_sample(
         self,
         noise,
@@ -45,9 +51,11 @@ class TrainGuider(comfy_extras.nodes_custom_sampler.Guider_Basic):
                 noise.shape,
                 self.conds,
                 self.model_options,
-                force_full_load=True,  # mirror behavior in TrainLoraNode.execute() to keep model loaded
+                force_full_load=not self.offloading,
+                force_offload=self.offloading,
             )
         )
+        torch.cuda.empty_cache()
         device = self.model_patcher.load_device
 
         if denoise_mask is not None:
@@ -404,16 +412,97 @@ def find_all_highest_child_module_with_forward(
     return result
 
 
-def patch(m):
+def find_modules_at_depth(
+    model: nn.Module, depth: int = 1, result=None, current_depth=0, name=None
+) -> list[nn.Module]:
+    """
+    Find modules at a specific depth level for gradient checkpointing.
+
+    Args:
+        model: The model to search
+        depth: Target depth level (1 = top-level blocks, 2 = their children, etc.)
+        result: Accumulator for results
+        current_depth: Current recursion depth
+        name: Current module name for logging
+
+    Returns:
+        List of modules at the target depth
+    """
+    if result is None:
+        result = []
+    name = name or "root"
+
+    # Skip container modules (they don't have meaningful forward)
+    is_container = isinstance(model, (nn.ModuleList, nn.Sequential, nn.ModuleDict))
+    has_forward = hasattr(model, "forward") and not is_container
+
+    if has_forward:
+        current_depth += 1
+        if current_depth == depth:
+            result.append(model)
+            logging.debug(f"Found module at depth {depth}: {name} ({model.__class__.__name__})")
+            return result
+
+    # Recurse into children
+    for next_name, child in model.named_children():
+        find_modules_at_depth(child, depth, result, current_depth, f"{name}.{next_name}")
+
+    return result
+
+
+class OffloadCheckpointFunction(torch.autograd.Function):
+    """
+    Gradient checkpointing that works with weight offloading.
+
+    Forward: no_grad -> compute -> weights can be freed
+    Backward: enable_grad -> recompute -> backward -> weights can be freed
+
+    For single input, single output modules (Linear, Conv*).
+    """
+
+    @staticmethod
+    def forward(ctx, x: torch.Tensor, forward_fn):
+        ctx.save_for_backward(x)
+        ctx.forward_fn = forward_fn
+        with torch.no_grad():
+            return forward_fn(x)
+
+    @staticmethod
+    def backward(ctx, grad_out: torch.Tensor):
+        x, = ctx.saved_tensors
+        forward_fn = ctx.forward_fn
+
+        # Clear context early
+        ctx.forward_fn = None
+
+        with torch.enable_grad():
+            x_detached = x.detach().requires_grad_(True)
+            y = forward_fn(x_detached)
+            y.backward(grad_out)
+            grad_x = x_detached.grad
+
+        # Explicit cleanup
+        del y, x_detached, forward_fn
+
+        return grad_x, None
+
+
+def patch(m, offloading=False):
     if not hasattr(m, "forward"):
         return
     org_forward = m.forward
 
-    def fwd(args, kwargs):
-        return org_forward(*args, **kwargs)
+    # Branch 1: Linear/Conv* -> offload-compatible checkpoint (single input/output)
+    if offloading and isinstance(m, (nn.Linear, nn.Conv1d, nn.Conv2d, nn.Conv3d)):
+        def checkpointing_fwd(x):
+            return OffloadCheckpointFunction.apply(x, org_forward)
+    # Branch 2: Others -> standard checkpoint
+    else:
+        def fwd(args, kwargs):
+            return org_forward(*args, **kwargs)
 
-    def checkpointing_fwd(*args, **kwargs):
-        return torch.utils.checkpoint.checkpoint(fwd, args, kwargs, use_reentrant=False)
+        def checkpointing_fwd(*args, **kwargs):
+            return torch.utils.checkpoint.checkpoint(fwd, args, kwargs, use_reentrant=False)
 
     m.org_forward = org_forward
     m.forward = checkpointing_fwd
@@ -936,6 +1025,18 @@ class TrainLoraNode(io.ComfyNode):
                     default=True,
                     tooltip="Use gradient checkpointing for training.",
                 ),
+                io.Int.Input(
+                    "checkpoint_depth",
+                    default=1,
+                    min=1,
+                    max=5,
+                    tooltip="Depth level for gradient checkpointing.",
+                ),
+                io.Boolean.Input(
+                    "offloading",
+                    default=False,
+                    tooltip="Depth level for gradient checkpointing.",
+                ),
                 io.Combo.Input(
                     "existing_lora",
                     options=folder_paths.get_filename_list("loras") + ["[None]"],
@@ -982,6 +1083,8 @@ class TrainLoraNode(io.ComfyNode):
         lora_dtype,
         algorithm,
         gradient_checkpointing,
+        checkpoint_depth,
+        offloading,
         existing_lora,
         bucket_mode,
         bypass_mode,
@@ -1000,6 +1103,8 @@ class TrainLoraNode(io.ComfyNode):
         lora_dtype = lora_dtype[0]
         algorithm = algorithm[0]
         gradient_checkpointing = gradient_checkpointing[0]
+        offloading = offloading[0]
+        checkpoint_depth = checkpoint_depth[0]
         existing_lora = existing_lora[0]
         bucket_mode = bucket_mode[0]
         bypass_mode = bypass_mode[0]
@@ -1054,16 +1159,18 @@ class TrainLoraNode(io.ComfyNode):
 
             # Setup gradient checkpointing
             if gradient_checkpointing:
-                for m in find_all_highest_child_module_with_forward(
-                    mp.model.diffusion_model
-                ):
-                    patch(m)
+                modules_to_patch = find_modules_at_depth(
+                    mp.model.diffusion_model, depth=checkpoint_depth
+                )
+                logging.info(f"Gradient checkpointing: patching {len(modules_to_patch)} modules at depth {checkpoint_depth}")
+                for m in modules_to_patch:
+                    patch(m, offloading=offloading)
 
             torch.cuda.empty_cache()
             # With force_full_load=False we should be able to have offloading
             # But for offloading in training we need custom AutoGrad hooks for fwd/bwd
             comfy.model_management.load_models_gpu(
-                [mp], memory_required=1e20, force_full_load=True
+                [mp], memory_required=1e20, force_full_load=not offloading
             )
             torch.cuda.empty_cache()
 
@@ -1100,7 +1207,7 @@ class TrainLoraNode(io.ComfyNode):
                 )
 
             # Setup guider
-            guider = TrainGuider(mp)
+            guider = TrainGuider(mp, offloading=offloading)
             guider.set_conds(positive)
 
             # Inject bypass hooks if bypass mode is enabled
@@ -1113,6 +1220,7 @@ class TrainLoraNode(io.ComfyNode):
 
             # Run training loop
             try:
+                comfy.model_management.in_training = True
                 _run_training_loop(
                     guider,
                     train_sampler,
@@ -1123,6 +1231,7 @@ class TrainLoraNode(io.ComfyNode):
                     multi_res,
                 )
             finally:
+                comfy.model_management.in_training = False
                 # Eject bypass hooks if they were injected
                 if bypass_injections is not None:
                     for injection in bypass_injections:
@@ -1132,19 +1241,20 @@ class TrainLoraNode(io.ComfyNode):
                     unpatch(m)
             del train_sampler, optimizer
 
-            # Finalize adapters
+            for param in lora_sd:
+                lora_sd[param] = lora_sd[param].to(lora_dtype).detach()
+
             for adapter in all_weight_adapters:
                 adapter.requires_grad_(False)
-
-            for param in lora_sd:
-                lora_sd[param] = lora_sd[param].to(lora_dtype)
+                del adapter
+            del all_weight_adapters
 
             # mp in train node is highly specialized for training
             # use it in inference will result in bad behavior so we don't return it
             return io.NodeOutput(lora_sd, loss_map, steps + existing_steps)
 
 
-class LoraModelLoader(io.ComfyNode):#
+class LoraModelLoader(io.ComfyNode):
     @classmethod
     def define_schema(cls):
         return io.Schema(
@@ -1166,6 +1276,11 @@ class LoraModelLoader(io.ComfyNode):#
                     max=100.0,
                     tooltip="How strongly to modify the diffusion model. This value can be negative.",
                 ),
+                io.Boolean.Input(
+                    "bypass",
+                    default=False,
+                    tooltip="When enabled, applies LoRA in bypass mode without modifying base model weights. Useful for training and when model weights are offloaded.",
+                ),
             ],
             outputs=[
                 io.Model.Output(
@@ -1175,13 +1290,18 @@ class LoraModelLoader(io.ComfyNode):#
         )
 
     @classmethod
-    def execute(cls, model, lora, strength_model):
+    def execute(cls, model, lora, strength_model, bypass=False):
         if strength_model == 0:
             return io.NodeOutput(model)
 
-        model_lora, _ = comfy.sd.load_lora_for_models(
-            model, None, lora, strength_model, 0
-        )
+        if bypass:
+            model_lora, _ = comfy.sd.load_bypass_lora_for_models(
+                model, None, lora, strength_model, 0
+            )
+        else:
+            model_lora, _ = comfy.sd.load_lora_for_models(
+                model, None, lora, strength_model, 0
+            )
         return io.NodeOutput(model_lora)
 
 

comfy_extras/nodes_video.py

@@ -202,6 +202,56 @@ class LoadVideo(io.ComfyNode):
 
         return True
 
+class VideoSlice(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="Video Slice",
+            display_name="Video Slice",
+            search_aliases=[
+                "trim video duration",
+                "skip first frames",
+                "frame load cap",
+                "start time",
+            ],
+            category="image/video",
+            inputs=[
+                io.Video.Input("video"),
+                io.Float.Input(
+                    "start_time",
+                    default=0.0,
+                    max=1e5,
+                    min=-1e5,
+                    step=0.001,
+                    tooltip="Start time in seconds",
+                ),
+                io.Float.Input(
+                    "duration",
+                    default=0.0,
+                    min=0.0,
+                    step=0.001,
+                    tooltip="Duration in seconds, or 0 for unlimited duration",
+                ),
+                io.Boolean.Input(
+                    "strict_duration",
+                    default=False,
+                    tooltip="If True, when the specified duration is not possible, an error will be raised.",
+                ),
+            ],
+            outputs=[
+                io.Video.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, video: io.Video.Type, start_time: float, duration: float, strict_duration: bool) -> io.NodeOutput:
+        trimmed = video.as_trimmed(start_time, duration, strict_duration=strict_duration)
+        if trimmed is not None:
+            return io.NodeOutput(trimmed)
+        raise ValueError(
+            f"Failed to slice video:\nSource duration: {video.get_duration()}\nStart time: {start_time}\nTarget duration: {duration}"
+        )
+
 
 class VideoExtension(ComfyExtension):
     @override
@@ -212,6 +262,7 @@ class VideoExtension(ComfyExtension):
             CreateVideo,
             GetVideoComponents,
             LoadVideo,
+            VideoSlice,
         ]
 
 async def comfy_entrypoint() -> VideoExtension:

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.12.0"
+__version__ = "0.13.0"

execution.py

@@ -13,8 +13,11 @@ from contextlib import nullcontext
 
 import torch
 
+from comfy.cli_args import args
 import comfy.memory_management
 import comfy.model_management
+import comfy_aimdo.model_vbar
+
 from latent_preview import set_preview_method
 import nodes
 from comfy_execution.caching import (
@@ -527,8 +530,10 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                     output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                 finally:
                     if allocator is not None:
+                        if args.verbose == "DEBUG":
+                            comfy_aimdo.model_vbar.vbars_analyze()
                         comfy.model_management.reset_cast_buffers()
-                        torch.cuda.synchronize()
+                        comfy_aimdo.model_vbar.vbars_reset_watermark_limits()
 
             if has_pending_tasks:
                 pending_async_nodes[unique_id] = output_data
@@ -975,6 +980,10 @@ async def validate_inputs(prompt_id, prompt, item, validated):
                 input_filtered[x] = input_data_all[x]
         if 'input_types' in validate_function_inputs:
             input_filtered['input_types'] = [received_types]
+        for x in list(input_filtered.keys()):
+            if input_filtered[x] is io.EmptyInputSentinel:
+                del input_filtered[x]
+
 
         ret = await _async_map_node_over_list(prompt_id, unique_id, obj_class, input_filtered, validate_function_name, v3_data=v3_data)
         ret = await resolve_map_node_over_list_results(ret)

main.py

@@ -192,7 +192,10 @@ import comfy_aimdo.control
 import comfy_aimdo.torch
 
 if enables_dynamic_vram():
-    if comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index):
+    if comfy.model_management.torch_version_numeric < (2, 8):
+        logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
+        comfy.memory_management.aimdo_allocator = None
+    elif comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index):
         if args.verbose == 'DEBUG':
             comfy_aimdo.control.set_log_debug()
         elif args.verbose == 'CRITICAL':
@@ -208,7 +211,7 @@ if enables_dynamic_vram():
         comfy.memory_management.aimdo_allocator = comfy_aimdo.torch.get_torch_allocator()
         logging.info("DynamicVRAM support detected and enabled")
     else:
-        logging.info("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
+        logging.warning("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
         comfy.memory_management.aimdo_allocator = None
 
 

nodes.py

@@ -2433,7 +2433,8 @@ async def init_builtin_extra_nodes():
         "nodes_image_compare.py",
         "nodes_zimage.py",
         "nodes_lora_debug.py",
-        "nodes_color.py"
+        "nodes_color.py",
+        "nodes_toolkit.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.12.0"
+version = "0.13.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,6 +1,6 @@
-comfyui-frontend-package==1.37.11
-comfyui-workflow-templates==0.8.31
-comfyui-embedded-docs==0.4.0
+comfyui-frontend-package==1.38.13
+comfyui-workflow-templates==0.8.38
+comfyui-embedded-docs==0.4.1
 torch
 torchsde
 torchvision
@@ -22,7 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
-comfy-aimdo>=0.1.7
+comfy-aimdo>=0.1.8
 requests
 
 #non essential dependencies:

tests/execution/test_jobs.py

@@ -5,8 +5,11 @@ from comfy_execution.jobs import (
     is_previewable,
     normalize_queue_item,
     normalize_history_item,
+    normalize_output_item,
+    normalize_outputs,
     get_outputs_summary,
     apply_sorting,
+    has_3d_extension,
 )
 
 
@@ -35,8 +38,8 @@ class TestIsPreviewable:
     """Unit tests for is_previewable()"""
 
     def test_previewable_media_types(self):
-        """Images, video, audio media types should be previewable."""
-        for media_type in ['images', 'video', 'audio']:
+        """Images, video, audio, 3d media types should be previewable."""
+        for media_type in ['images', 'video', 'audio', '3d']:
             assert is_previewable(media_type, {}) is True
 
     def test_non_previewable_media_types(self):
@@ -46,7 +49,7 @@ class TestIsPreviewable:
 
     def test_3d_extensions_previewable(self):
         """3D file extensions should be previewable regardless of media_type."""
-        for ext in ['.obj', '.fbx', '.gltf', '.glb']:
+        for ext in ['.obj', '.fbx', '.gltf', '.glb', '.usdz']:
             item = {'filename': f'model{ext}'}
             assert is_previewable('files', item) is True
 
@@ -160,7 +163,7 @@ class TestGetOutputsSummary:
 
     def test_3d_files_previewable(self):
         """3D file extensions should be previewable."""
-        for ext in ['.obj', '.fbx', '.gltf', '.glb']:
+        for ext in ['.obj', '.fbx', '.gltf', '.glb', '.usdz']:
             outputs = {
                 'node1': {
                     'files': [{'filename': f'model{ext}', 'type': 'output'}]
@@ -192,6 +195,64 @@ class TestGetOutputsSummary:
         assert preview['mediaType'] == 'images'
         assert preview['subfolder'] == 'outputs'
 
+    def test_string_3d_filename_creates_preview(self):
+        """String items with 3D extensions should synthesize a preview (Preview3D node output).
+        Only the .glb counts — nulls and non-file strings are excluded."""
+        outputs = {
+            'node1': {
+                'result': ['preview3d_abc123.glb', None, None]
+            }
+        }
+        count, preview = get_outputs_summary(outputs)
+        assert count == 1
+        assert preview is not None
+        assert preview['filename'] == 'preview3d_abc123.glb'
+        assert preview['mediaType'] == '3d'
+        assert preview['nodeId'] == 'node1'
+        assert preview['type'] == 'output'
+
+    def test_string_non_3d_filename_no_preview(self):
+        """String items without 3D extensions should not create a preview."""
+        outputs = {
+            'node1': {
+                'result': ['data.json', None]
+            }
+        }
+        count, preview = get_outputs_summary(outputs)
+        assert count == 0
+        assert preview is None
+
+    def test_string_3d_filename_used_as_fallback(self):
+        """String 3D preview should be used when no dict items are previewable."""
+        outputs = {
+            'node1': {
+                'latents': [{'filename': 'latent.safetensors'}],
+            },
+            'node2': {
+                'result': ['model.glb', None]
+            }
+        }
+        count, preview = get_outputs_summary(outputs)
+        assert preview is not None
+        assert preview['filename'] == 'model.glb'
+        assert preview['mediaType'] == '3d'
+
+
+class TestHas3DExtension:
+    """Unit tests for has_3d_extension()"""
+
+    def test_recognized_extensions(self):
+        for ext in ['.obj', '.fbx', '.gltf', '.glb', '.usdz']:
+            assert has_3d_extension(f'model{ext}') is True
+
+    def test_case_insensitive(self):
+        assert has_3d_extension('MODEL.GLB') is True
+        assert has_3d_extension('Scene.GLTF') is True
+
+    def test_non_3d_extensions(self):
+        for name in ['photo.png', 'video.mp4', 'data.json', 'model']:
+            assert has_3d_extension(name) is False
+
 
 class TestApplySorting:
     """Unit tests for apply_sorting()"""
@@ -395,3 +456,142 @@ class TestNormalizeHistoryItem:
             'prompt': {'nodes': {'1': {}}},
             'extra_data': {'create_time': 1234567890, 'client_id': 'abc'},
         }
+
+    def test_include_outputs_normalizes_3d_strings(self):
+        """Detail view should transform string 3D filenames into file output dicts."""
+        history_item = {
+            'prompt': (
+                5,
+                'prompt-3d',
+                {'nodes': {}},
+                {'create_time': 1234567890},
+                ['node1'],
+            ),
+            'status': {'status_str': 'success', 'completed': True, 'messages': []},
+            'outputs': {
+                'node1': {
+                    'result': ['preview3d_abc123.glb', None, None]
+                }
+            },
+        }
+        job = normalize_history_item('prompt-3d', history_item, include_outputs=True)
+
+        assert job['outputs_count'] == 1
+        result_items = job['outputs']['node1']['result']
+        assert len(result_items) == 1
+        assert result_items[0] == {
+            'filename': 'preview3d_abc123.glb',
+            'type': 'output',
+            'subfolder': '',
+            'mediaType': '3d',
+        }
+
+    def test_include_outputs_preserves_dict_items(self):
+        """Detail view normalization should pass dict items through unchanged."""
+        history_item = {
+            'prompt': (
+                5,
+                'prompt-img',
+                {'nodes': {}},
+                {'create_time': 1234567890},
+                ['node1'],
+            ),
+            'status': {'status_str': 'success', 'completed': True, 'messages': []},
+            'outputs': {
+                'node1': {
+                    'images': [
+                        {'filename': 'photo.png', 'type': 'output', 'subfolder': ''},
+                    ]
+                }
+            },
+        }
+        job = normalize_history_item('prompt-img', history_item, include_outputs=True)
+
+        assert job['outputs_count'] == 1
+        assert job['outputs']['node1']['images'] == [
+            {'filename': 'photo.png', 'type': 'output', 'subfolder': ''},
+        ]
+
+
+class TestNormalizeOutputItem:
+    """Unit tests for normalize_output_item()"""
+
+    def test_none_returns_none(self):
+        assert normalize_output_item(None) is None
+
+    def test_string_3d_extension_synthesizes_dict(self):
+        result = normalize_output_item('model.glb')
+        assert result == {'filename': 'model.glb', 'type': 'output', 'subfolder': '', 'mediaType': '3d'}
+
+    def test_string_non_3d_extension_returns_none(self):
+        assert normalize_output_item('data.json') is None
+
+    def test_string_no_extension_returns_none(self):
+        assert normalize_output_item('camera_info_string') is None
+
+    def test_dict_passes_through(self):
+        item = {'filename': 'test.png', 'type': 'output'}
+        assert normalize_output_item(item) is item
+
+    def test_other_types_return_none(self):
+        assert normalize_output_item(42) is None
+        assert normalize_output_item(True) is None
+
+
+class TestNormalizeOutputs:
+    """Unit tests for normalize_outputs()"""
+
+    def test_empty_outputs(self):
+        assert normalize_outputs({}) == {}
+
+    def test_dict_items_pass_through(self):
+        outputs = {
+            'node1': {
+                'images': [{'filename': 'a.png', 'type': 'output'}],
+            }
+        }
+        result = normalize_outputs(outputs)
+        assert result == outputs
+
+    def test_3d_string_synthesized(self):
+        outputs = {
+            'node1': {
+                'result': ['model.glb', None, None],
+            }
+        }
+        result = normalize_outputs(outputs)
+        assert result == {
+            'node1': {
+                'result': [
+                    {'filename': 'model.glb', 'type': 'output', 'subfolder': '', 'mediaType': '3d'},
+                ],
+            }
+        }
+
+    def test_animated_key_preserved(self):
+        outputs = {
+            'node1': {
+                'images': [{'filename': 'a.png', 'type': 'output'}],
+                'animated': [True],
+            }
+        }
+        result = normalize_outputs(outputs)
+        assert result['node1']['animated'] == [True]
+
+    def test_non_dict_node_outputs_preserved(self):
+        outputs = {'node1': 'unexpected_value'}
+        result = normalize_outputs(outputs)
+        assert result == {'node1': 'unexpected_value'}
+
+    def test_none_items_filtered_but_other_types_preserved(self):
+        outputs = {
+            'node1': {
+                'result': ['data.json', None, [1, 2, 3]],
+            }
+        }
+        result = normalize_outputs(outputs)
+        assert result == {
+            'node1': {
+                'result': ['data.json', [1, 2, 3]],
+            }
+        }