Basic implementation of z image fun control union 2.0 (#11304)

The inpaint part is currently missing and will be implemented later.

I think they messed up this model pretty bad. They added some
control_noise_refiner blocks but don't actually use them. There is a typo
in their code so instead of doing control_noise_refiner -> control_layers
it runs the whole control_layers twice.

Unfortunately they trained with this typo so the model works but is kind
of slow and would probably perform a lot better if they corrected their
code and trained it again.

comfy/ldm/lumina/controlnet.py

@@ -41,6 +41,11 @@ class ZImage_Control(torch.nn.Module):
         ffn_dim_multiplier: float = (8.0 / 3.0),
         norm_eps: float = 1e-5,
         qk_norm: bool = True,
+        n_control_layers=6,
+        control_in_dim=16,
+        additional_in_dim=0,
+        broken=False,
+        refiner_control=False,
         dtype=None,
         device=None,
         operations=None,
@@ -49,10 +54,11 @@ class ZImage_Control(torch.nn.Module):
         super().__init__()
         operation_settings = {"operations": operations, "device": device, "dtype": dtype}
 
-        self.additional_in_dim = 0
-        self.control_in_dim = 16
+        self.broken = broken
+        self.additional_in_dim = additional_in_dim
+        self.control_in_dim = control_in_dim
         n_refiner_layers = 2
-        self.n_control_layers = 6
+        self.n_control_layers = n_control_layers
         self.control_layers = nn.ModuleList(
             [
                 ZImageControlTransformerBlock(
@@ -74,28 +80,49 @@ class ZImage_Control(torch.nn.Module):
         all_x_embedder = {}
         patch_size = 2
         f_patch_size = 1
-        x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * self.control_in_dim, dim, bias=True, device=device, dtype=dtype)
+        x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * (self.control_in_dim + self.additional_in_dim), dim, bias=True, device=device, dtype=dtype)
         all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder
 
+        self.refiner_control = refiner_control
+
         self.control_all_x_embedder = nn.ModuleDict(all_x_embedder)
-        self.control_noise_refiner = nn.ModuleList(
-            [
-                JointTransformerBlock(
-                    layer_id,
-                    dim,
-                    n_heads,
-                    n_kv_heads,
-                    multiple_of,
-                    ffn_dim_multiplier,
-                    norm_eps,
-                    qk_norm,
-                    modulation=True,
-                    z_image_modulation=True,
-                    operation_settings=operation_settings,
-                )
-                for layer_id in range(n_refiner_layers)
-            ]
-        )
+        if self.refiner_control:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    ZImageControlTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        block_id=layer_id,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+        else:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    JointTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        modulation=True,
+                        z_image_modulation=True,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
 
     def forward(self, cap_feats, control_context, x_freqs_cis, adaln_input):
         patch_size = 2
@@ -105,9 +132,29 @@ class ZImage_Control(torch.nn.Module):
         control_context = self.control_all_x_embedder[f"{patch_size}-{f_patch_size}"](control_context.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
 
         x_attn_mask = None
-        for layer in self.control_noise_refiner:
-            control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input)
+        if not self.refiner_control:
+            for layer in self.control_noise_refiner:
+                control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input)
+
         return control_context
 
+    def forward_noise_refiner_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input):
+        if self.refiner_control:
+            if self.broken:
+                if layer_id == 0:
+                    return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+                if layer_id > 0:
+                    out = None
+                    for i in range(1, len(self.control_layers)):
+                        o, control_context = self.control_layers[i](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+                        if out is None:
+                            out = o
+
+                    return (out, control_context)
+            else:
+                return self.control_noise_refiner[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+        else:
+            return (None, control_context)
+
     def forward_control_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input):
         return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)

comfy/ldm/lumina/model.py

@@ -536,6 +536,7 @@ class NextDiT(nn.Module):
         bsz = len(x)
         pH = pW = self.patch_size
         device = x[0].device
+        orig_x = x
 
         if self.pad_tokens_multiple is not None:
             pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple
@@ -572,13 +573,21 @@ class NextDiT(nn.Module):
 
         freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)
 
+        patches = transformer_options.get("patches", {})
+
         # refine context
         for layer in self.context_refiner:
             cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options)
 
         padded_img_mask = None
-        for layer in self.noise_refiner:
+        x_input = x
+        for i, layer in enumerate(self.noise_refiner):
             x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options)
+            if "noise_refiner" in patches:
+                for p in patches["noise_refiner"]:
+                    out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": freqs_cis[:, cap_pos_ids.shape[1]:], "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"})
+                    if "img" in out:
+                        x = out["img"]
 
         padded_full_embed = torch.cat((cap_feats, x), dim=1)
         mask = None
@@ -622,14 +631,15 @@ class NextDiT(nn.Module):
 
         patches = transformer_options.get("patches", {})
         x_is_tensor = isinstance(x, torch.Tensor)
-        img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options)
+        img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options)
         freqs_cis = freqs_cis.to(img.device)
 
+        img_input = img
         for i, layer in enumerate(self.layers):
             img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
             if "double_block" in patches:
                 for p in patches["double_block"]:
-                    out = p({"img": img[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options})
+                    out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options})
                     if "img" in out:
                         img[:, cap_size[0]:] = out["img"]
                     if "txt" in out: