Revert "Fix some fp8 scaled checkpoints no longer working. (#13239)"

This reverts commit 4504ee718a.

comfy/ldm/flux/model.py

@@ -386,7 +386,7 @@ class Flux(nn.Module):
                     h = max(h, ref.shape[-2] + h_offset)
                     w = max(w, ref.shape[-1] + w_offset)
 
-                kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset, transformer_options=transformer_options)
+                kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
                 img = torch.cat([img, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
                 ref_num_tokens.append(kontext.shape[1])

comfy/ldm/lightricks/av_model.py

@@ -681,33 +681,6 @@ class LTXAVModel(LTXVModel):
         additional_args["has_spatial_mask"] = has_spatial_mask
 
         ax, a_latent_coords = self.a_patchifier.patchify(ax)
-
-        # Inject reference audio for ID-LoRA in-context conditioning
-        ref_audio = kwargs.get("ref_audio", None)
-        ref_audio_seq_len = 0
-        if ref_audio is not None:
-            ref_tokens = ref_audio["tokens"].to(dtype=ax.dtype, device=ax.device)
-            if ref_tokens.shape[0] < ax.shape[0]:
-                ref_tokens = ref_tokens.expand(ax.shape[0], -1, -1)
-            ref_audio_seq_len = ref_tokens.shape[1]
-            B = ax.shape[0]
-
-            # Compute negative temporal positions matching ID-LoRA convention:
-            # offset by -(end_of_last_token + time_per_latent) so reference ends just before t=0
-            p = self.a_patchifier
-            tpl = p.hop_length * p.audio_latent_downsample_factor / p.sample_rate
-            ref_start = p._get_audio_latent_time_in_sec(0, ref_audio_seq_len, torch.float32, ax.device)
-            ref_end = p._get_audio_latent_time_in_sec(1, ref_audio_seq_len + 1, torch.float32, ax.device)
-            time_offset = ref_end[-1].item() + tpl
-            ref_start = (ref_start - time_offset).unsqueeze(0).expand(B, -1).unsqueeze(1)
-            ref_end = (ref_end - time_offset).unsqueeze(0).expand(B, -1).unsqueeze(1)
-            ref_pos = torch.stack([ref_start, ref_end], dim=-1)
-
-            additional_args["ref_audio_seq_len"] = ref_audio_seq_len
-            additional_args["target_audio_seq_len"] = ax.shape[1]
-            ax = torch.cat([ref_tokens, ax], dim=1)
-            a_latent_coords = torch.cat([ref_pos.to(a_latent_coords), a_latent_coords], dim=2)
-
         ax = self.audio_patchify_proj(ax)
 
         # additional_args.update({"av_orig_shape": list(x.shape)})
@@ -748,14 +721,6 @@ class LTXAVModel(LTXVModel):
 
         # Prepare audio timestep
         a_timestep = kwargs.get("a_timestep")
-        ref_audio_seq_len = kwargs.get("ref_audio_seq_len", 0)
-        if ref_audio_seq_len > 0 and a_timestep is not None:
-            # Reference tokens must have timestep=0, expand scalar/1D timestep to per-token so ref=0 and target=sigma.
-            target_len = kwargs.get("target_audio_seq_len")
-            if a_timestep.dim() <= 1:
-                a_timestep = a_timestep.view(-1, 1).expand(batch_size, target_len)
-            ref_ts = torch.zeros(batch_size, ref_audio_seq_len, *a_timestep.shape[2:], device=a_timestep.device, dtype=a_timestep.dtype)
-            a_timestep = torch.cat([ref_ts, a_timestep], dim=1)
         if a_timestep is not None:
             a_timestep_scaled = a_timestep * self.timestep_scale_multiplier
             a_timestep_flat = a_timestep_scaled.flatten()
@@ -990,13 +955,6 @@ class LTXAVModel(LTXVModel):
         v_embedded_timestep = embedded_timestep[0]
         a_embedded_timestep = embedded_timestep[1]
 
-        # Trim reference audio tokens before unpatchification
-        ref_audio_seq_len = kwargs.get("ref_audio_seq_len", 0)
-        if ref_audio_seq_len > 0:
-            ax = ax[:, ref_audio_seq_len:]
-            if a_embedded_timestep.shape[1] > 1:
-                a_embedded_timestep = a_embedded_timestep[:, ref_audio_seq_len:]
-
         # Expand compressed video timestep if needed
         if isinstance(v_embedded_timestep, CompressedTimestep):
             v_embedded_timestep = v_embedded_timestep.expand()

comfy/model_base.py

@@ -937,10 +937,9 @@ class LongCatImage(Flux):
         transformer_options = transformer_options.copy()
         rope_opts = transformer_options.get("rope_options", {})
         rope_opts = dict(rope_opts)
-        pe_len = float(c_crossattn.shape[1]) if c_crossattn is not None else 512.0
         rope_opts.setdefault("shift_t", 1.0)
-        rope_opts.setdefault("shift_y", pe_len)
-        rope_opts.setdefault("shift_x", pe_len)
+        rope_opts.setdefault("shift_y", 512.0)
+        rope_opts.setdefault("shift_x", 512.0)
         transformer_options["rope_options"] = rope_opts
         return super()._apply_model(x, t, c_concat, c_crossattn, control, transformer_options, **kwargs)
 
@@ -1061,10 +1060,6 @@ class LTXAV(BaseModel):
         if guide_attention_entries is not None:
             out['guide_attention_entries'] = comfy.conds.CONDConstant(guide_attention_entries)
 
-        ref_audio = kwargs.get("ref_audio", None)
-        if ref_audio is not None:
-            out['ref_audio'] = comfy.conds.CONDConstant(ref_audio)
-
         return out
 
     def process_timestep(self, timestep, x, denoise_mask=None, audio_denoise_mask=None, **kwargs):

comfy/model_management.py

@@ -1325,9 +1325,9 @@ MAX_PINNED_MEMORY = -1
 if not args.disable_pinned_memory:
     if is_nvidia() or is_amd():
         if WINDOWS:
-            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45  # Windows limit is apparently 50%
+            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.40  # Windows limit is apparently 50%
         else:
-            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95
+            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.90
         logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024)))
 
 PINNING_ALLOWED_TYPES = set(["Tensor", "Parameter", "QuantizedTensor"])
@@ -1402,8 +1402,6 @@ def unpin_memory(tensor):
 
     if torch.cuda.cudart().cudaHostUnregister(ptr) == 0:
         TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr)
-        if len(PINNED_MEMORY) == 0:
-            TOTAL_PINNED_MEMORY = 0
         return True
     else:
         logging.warning("Unpin error.")

comfy/text_encoders/llama.py

@@ -1028,19 +1028,12 @@ class Qwen25_7BVLI(BaseLlama, BaseGenerate, torch.nn.Module):
                 grid = e.get("extra", None)
                 start = e.get("index")
                 if position_ids is None:
-                    position_ids = torch.ones((3, embeds.shape[1]), device=embeds.device, dtype=torch.long)
+                    position_ids = torch.zeros((3, embeds.shape[1]), device=embeds.device)
                     position_ids[:, :start] = torch.arange(0, start, device=embeds.device)
                 end = e.get("size") + start
                 len_max = int(grid.max()) // 2
                 start_next = len_max + start
-                if attention_mask is not None:
-                    # Assign compact sequential positions to attended tokens only,
-                    # skipping over padding so post-padding tokens aren't inflated.
-                    after_mask = attention_mask[0, end:]
-                    text_positions = after_mask.cumsum(0) - 1 + start_next + offset
-                    position_ids[:, end:] = torch.where(after_mask.bool(), text_positions, position_ids[0, end:])
-                else:
-                    position_ids[:, end:] = torch.arange(start_next + offset, start_next + (embeds.shape[1] - end) + offset, device=embeds.device)
+                position_ids[:, end:] = torch.arange(start_next + offset, start_next + (embeds.shape[1] - end) + offset, device=embeds.device)
                 position_ids[0, start:end] = start + offset
                 max_d = int(grid[0][1]) // 2
                 position_ids[1, start:end] = torch.arange(start + offset, start + max_d + offset, device=embeds.device).unsqueeze(1).repeat(1, math.ceil((end - start) / max_d)).flatten(0)[:end - start]

comfy/text_encoders/longcat_image.py

@@ -64,13 +64,7 @@ class LongCatImageBaseTokenizer(Qwen25_7BVLITokenizer):
         return [output]
 
 
-IMAGE_PAD_TOKEN_ID = 151655
-
 class LongCatImageTokenizer(sd1_clip.SD1Tokenizer):
-    T2I_PREFIX = "<|im_start|>system\nAs an image captioning expert, generate a descriptive text prompt based on an image content, suitable for input to a text-to-image model.<|im_end|>\n<|im_start|>user\n"
-    EDIT_PREFIX = "<|im_start|>system\nAs an image editing expert, first analyze the content and attributes of the input image(s). Then, based on the user's editing instructions, clearly and precisely determine how to modify the given image(s), ensuring that only the specified parts are altered and all other aspects remain consistent with the original(s).<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>"
-    SUFFIX = "<|im_end|>\n<|im_start|>assistant\n"
-
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(
             embedding_directory=embedding_directory,
@@ -78,8 +72,10 @@ class LongCatImageTokenizer(sd1_clip.SD1Tokenizer):
             name="qwen25_7b",
             tokenizer=LongCatImageBaseTokenizer,
         )
+        self.longcat_template_prefix = "<|im_start|>system\nAs an image captioning expert, generate a descriptive text prompt based on an image content, suitable for input to a text-to-image model.<|im_end|>\n<|im_start|>user\n"
+        self.longcat_template_suffix = "<|im_end|>\n<|im_start|>assistant\n"
 
-    def tokenize_with_weights(self, text, return_word_ids=False, images=None, **kwargs):
+    def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
         skip_template = False
         if text.startswith("<|im_start|>"):
             skip_template = True
@@ -94,14 +90,11 @@ class LongCatImageTokenizer(sd1_clip.SD1Tokenizer):
                 text, return_word_ids=return_word_ids, disable_weights=True, **kwargs
             )
         else:
-            has_images = images is not None and len(images) > 0
-            template_prefix = self.EDIT_PREFIX if has_images else self.T2I_PREFIX
-
             prefix_ids = base_tok.tokenizer(
-                template_prefix, add_special_tokens=False
+                self.longcat_template_prefix, add_special_tokens=False
             )["input_ids"]
             suffix_ids = base_tok.tokenizer(
-                self.SUFFIX, add_special_tokens=False
+                self.longcat_template_suffix, add_special_tokens=False
             )["input_ids"]
 
             prompt_tokens = base_tok.tokenize_with_weights(
@@ -113,14 +106,6 @@ class LongCatImageTokenizer(sd1_clip.SD1Tokenizer):
             suffix_pairs = [(t, 1.0) for t in suffix_ids]
 
             combined = prefix_pairs + prompt_pairs + suffix_pairs
-
-            if has_images:
-                embed_count = 0
-                for i in range(len(combined)):
-                    if combined[i][0] == IMAGE_PAD_TOKEN_ID and embed_count < len(images):
-                        combined[i] = ({"type": "image", "data": images[embed_count], "original_type": "image"}, combined[i][1])
-                        embed_count += 1
-
             tokens = {"qwen25_7b": [combined]}
 
         return tokens

comfy/text_encoders/qwen_vl.py

@@ -425,7 +425,4 @@ class Qwen2VLVisionTransformer(nn.Module):
             hidden_states = block(hidden_states, position_embeddings, cu_seqlens_now, optimized_attention=optimized_attention)
 
         hidden_states = self.merger(hidden_states)
-        # Potentially important for spatially precise edits. This is present in the HF implementation.
-        reverse_indices = torch.argsort(window_index)
-        hidden_states = hidden_states[reverse_indices, :]
         return hidden_states

comfy_api_nodes/apis/wan.py

@@ -0,0 +1,226 @@
+from pydantic import BaseModel, Field
+
+
+class Text2ImageInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+
+
+class Image2ImageInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    images: list[str] = Field(..., min_length=1, max_length=2)
+
+
+class Text2VideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    audio_url: str | None = Field(None)
+
+
+class Image2VideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    img_url: str = Field(...)
+    audio_url: str | None = Field(None)
+
+
+class Reference2VideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    reference_video_urls: list[str] = Field(...)
+
+
+class Txt2ImageParametersField(BaseModel):
+    size: str = Field(...)
+    n: int = Field(1, description="Number of images to generate.")  # we support only value=1
+    seed: int = Field(..., ge=0, le=2147483647)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+
+
+class Image2ImageParametersField(BaseModel):
+    size: str | None = Field(None)
+    n: int = Field(1, description="Number of images to generate.")  # we support only value=1
+    seed: int = Field(..., ge=0, le=2147483647)
+    watermark: bool = Field(False)
+
+
+class Text2VideoParametersField(BaseModel):
+    size: str = Field(...)
+    seed: int = Field(..., ge=0, le=2147483647)
+    duration: int = Field(5, ge=5, le=15)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+    audio: bool = Field(False, description="Whether to generate audio automatically.")
+    shot_type: str = Field("single")
+
+
+class Image2VideoParametersField(BaseModel):
+    resolution: str = Field(...)
+    seed: int = Field(..., ge=0, le=2147483647)
+    duration: int = Field(5, ge=5, le=15)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+    audio: bool = Field(False, description="Whether to generate audio automatically.")
+    shot_type: str = Field("single")
+
+
+class Reference2VideoParametersField(BaseModel):
+    size: str = Field(...)
+    duration: int = Field(5, ge=5, le=15)
+    shot_type: str = Field("single")
+    seed: int = Field(..., ge=0, le=2147483647)
+    watermark: bool = Field(False)
+
+
+class Text2ImageTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Text2ImageInputField = Field(...)
+    parameters: Txt2ImageParametersField = Field(...)
+
+
+class Image2ImageTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Image2ImageInputField = Field(...)
+    parameters: Image2ImageParametersField = Field(...)
+
+
+class Text2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Text2VideoInputField = Field(...)
+    parameters: Text2VideoParametersField = Field(...)
+
+
+class Image2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Image2VideoInputField = Field(...)
+    parameters: Image2VideoParametersField = Field(...)
+
+
+class Reference2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Reference2VideoInputField = Field(...)
+    parameters: Reference2VideoParametersField = Field(...)
+
+
+class Wan27MediaItem(BaseModel):
+    type: str = Field(...)
+    url: str = Field(...)
+
+
+class Wan27ReferenceVideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    media: list[Wan27MediaItem] = Field(...)
+
+
+class Wan27ReferenceVideoParametersField(BaseModel):
+    resolution: str = Field(...)
+    ratio: str | None = Field(None)
+    duration: int = Field(5, ge=2, le=10)
+    watermark: bool = Field(False)
+    seed: int = Field(..., ge=0, le=2147483647)
+
+
+class Wan27ReferenceVideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Wan27ReferenceVideoInputField = Field(...)
+    parameters: Wan27ReferenceVideoParametersField = Field(...)
+
+
+class Wan27ImageToVideoInputField(BaseModel):
+    prompt: str | None = Field(None)
+    negative_prompt: str | None = Field(None)
+    media: list[Wan27MediaItem] = Field(...)
+
+
+class Wan27ImageToVideoParametersField(BaseModel):
+    resolution: str = Field(...)
+    duration: int = Field(5, ge=2, le=15)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+    seed: int = Field(..., ge=0, le=2147483647)
+
+
+class Wan27ImageToVideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Wan27ImageToVideoInputField = Field(...)
+    parameters: Wan27ImageToVideoParametersField = Field(...)
+
+
+class Wan27VideoEditInputField(BaseModel):
+    prompt: str = Field(...)
+    media: list[Wan27MediaItem] = Field(...)
+
+
+class Wan27VideoEditParametersField(BaseModel):
+    resolution: str = Field(...)
+    ratio: str | None = Field(None)
+    duration: int = Field(0)
+    audio_setting: str = Field("auto")
+    watermark: bool = Field(False)
+    seed: int = Field(..., ge=0, le=2147483647)
+
+
+class Wan27VideoEditTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Wan27VideoEditInputField = Field(...)
+    parameters: Wan27VideoEditParametersField = Field(...)
+
+
+class Wan27Text2VideoParametersField(BaseModel):
+    resolution: str = Field(...)
+    ratio: str | None = Field(None)
+    duration: int = Field(5, ge=2, le=15)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+    seed: int = Field(..., ge=0, le=2147483647)
+
+
+class Wan27Text2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Text2VideoInputField = Field(...)
+    parameters: Wan27Text2VideoParametersField = Field(...)
+
+
+class TaskCreationOutputField(BaseModel):
+    task_id: str = Field(...)
+    task_status: str = Field(...)
+
+
+class TaskCreationResponse(BaseModel):
+    output: TaskCreationOutputField | None = Field(None)
+    request_id: str = Field(...)
+    code: str | None = Field(None, description="Error code for the failed request.")
+    message: str | None = Field(None, description="Details about the failed request.")
+
+
+class TaskResult(BaseModel):
+    url: str | None = Field(None)
+    code: str | None = Field(None)
+    message: str | None = Field(None)
+
+
+class ImageTaskStatusOutputField(TaskCreationOutputField):
+    task_id: str = Field(...)
+    task_status: str = Field(...)
+    results: list[TaskResult] | None = Field(None)
+
+
+class VideoTaskStatusOutputField(TaskCreationOutputField):
+    task_id: str = Field(...)
+    task_status: str = Field(...)
+    video_url: str | None = Field(None)
+    code: str | None = Field(None)
+    message: str | None = Field(None)
+
+
+class ImageTaskStatusResponse(BaseModel):
+    output: ImageTaskStatusOutputField | None = Field(None)
+    request_id: str = Field(...)
+
+
+class VideoTaskStatusResponse(BaseModel):
+    output: VideoTaskStatusOutputField | None = Field(None)
+    request_id: str = Field(...)

comfy_api_nodes/nodes_topaz.py

@@ -38,6 +38,7 @@ from comfy_api_nodes.util import (
 UPSCALER_MODELS_MAP = {
     "Starlight (Astra) Fast": "slf-1",
     "Starlight (Astra) Creative": "slc-1",
+    "Starlight Precise 2.5": "slp-2.5",
 }
 
 

comfy_extras/nodes_lt.py

@@ -3,7 +3,6 @@ import node_helpers
 import torch
 import comfy.model_management
 import comfy.model_sampling
-import comfy.samplers
 import comfy.utils
 import math
 import numpy as np
@@ -683,84 +682,6 @@ class LTXVSeparateAVLatent(io.ComfyNode):
         return io.NodeOutput(video_latent, audio_latent)
 
 
-class LTXVReferenceAudio(io.ComfyNode):
-    @classmethod
-    def define_schema(cls) -> io.Schema:
-        return io.Schema(
-            node_id="LTXVReferenceAudio",
-            display_name="LTXV Reference Audio (ID-LoRA)",
-            category="conditioning/audio",
-            description="Set reference audio for ID-LoRA speaker identity transfer. Encodes a reference audio clip into the conditioning and optionally patches the model with identity guidance (extra forward pass without reference, amplifying the speaker identity effect).",
-            inputs=[
-                io.Model.Input("model"),
-                io.Conditioning.Input("positive"),
-                io.Conditioning.Input("negative"),
-                io.Audio.Input("reference_audio", tooltip="Reference audio clip whose speaker identity to transfer. ~5 seconds recommended (training duration). Shorter or longer clips may degrade voice identity transfer."),
-                io.Vae.Input(id="audio_vae", display_name="Audio VAE", tooltip="LTXV Audio VAE for encoding."),
-                io.Float.Input("identity_guidance_scale", default=3.0, min=0.0, max=100.0, step=0.01, round=0.01, tooltip="Strength of identity guidance. Runs an extra forward pass without reference each step to amplify speaker identity. Set to 0 to disable (no extra pass)."),
-                io.Float.Input("start_percent", default=0.0, min=0.0, max=1.0, step=0.001, advanced=True, tooltip="Start of the sigma range where identity guidance is active."),
-                io.Float.Input("end_percent", default=1.0, min=0.0, max=1.0, step=0.001, advanced=True, tooltip="End of the sigma range where identity guidance is active."),
-            ],
-            outputs=[
-                io.Model.Output(),
-                io.Conditioning.Output(display_name="positive"),
-                io.Conditioning.Output(display_name="negative"),
-            ],
-        )
-
-    @classmethod
-    def execute(cls, model, positive, negative, reference_audio, audio_vae, identity_guidance_scale, start_percent, end_percent) -> io.NodeOutput:
-        # Encode reference audio to latents and patchify
-        audio_latents = audio_vae.encode(reference_audio)
-        b, c, t, f = audio_latents.shape
-        ref_tokens = audio_latents.permute(0, 2, 1, 3).reshape(b, t, c * f)
-        ref_audio = {"tokens": ref_tokens}
-
-        positive = node_helpers.conditioning_set_values(positive, {"ref_audio": ref_audio})
-        negative = node_helpers.conditioning_set_values(negative, {"ref_audio": ref_audio})
-
-        # Patch model with identity guidance
-        m = model.clone()
-        scale = identity_guidance_scale
-        model_sampling = m.get_model_object("model_sampling")
-        sigma_start = model_sampling.percent_to_sigma(start_percent)
-        sigma_end = model_sampling.percent_to_sigma(end_percent)
-
-        def post_cfg_function(args):
-            if scale == 0:
-                return args["denoised"]
-
-            sigma = args["sigma"]
-            sigma_ = sigma[0].item()
-            if sigma_ > sigma_start or sigma_ < sigma_end:
-                return args["denoised"]
-
-            cond_pred = args["cond_denoised"]
-            cond = args["cond"]
-            cfg_result = args["denoised"]
-            model_options = args["model_options"].copy()
-            x = args["input"]
-
-            # Strip ref_audio from conditioning for the no-reference pass
-            noref_cond = []
-            for entry in cond:
-                new_entry = entry.copy()
-                mc = new_entry.get("model_conds", {}).copy()
-                mc.pop("ref_audio", None)
-                new_entry["model_conds"] = mc
-                noref_cond.append(new_entry)
-
-            (pred_noref,) = comfy.samplers.calc_cond_batch(
-                args["model"], [noref_cond], x, sigma, model_options
-            )
-
-            return cfg_result + (cond_pred - pred_noref) * scale
-
-        m.set_model_sampler_post_cfg_function(post_cfg_function)
-
-        return io.NodeOutput(m, positive, negative)
-
-
 class LtxvExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
@@ -776,7 +697,6 @@ class LtxvExtension(ComfyExtension):
             LTXVCropGuides,
             LTXVConcatAVLatent,
             LTXVSeparateAVLatent,
-            LTXVReferenceAudio,
         ]
 
 

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.18.1"
+__version__ = "0.18.4"

main.py

@@ -471,9 +471,6 @@ if __name__ == "__main__":
     if sys.version_info.major == 3 and sys.version_info.minor < 10:
         logging.warning("WARNING: You are using a python version older than 3.10, please upgrade to a newer one. 3.12 and above is recommended.")
 
-    if args.disable_dynamic_vram:
-        logging.warning("Dynamic vram disabled with argument. If you have any issues with dynamic vram enabled please give us a detailed reports as this argument will be removed soon.")
-
     event_loop, _, start_all_func = start_comfyui()
     try:
         x = start_all_func()

manager_requirements.txt

@@ -1 +1 @@
-comfyui_manager==4.1b8
+comfyui_manager==4.1b6

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.18.1"
+version = "0.18.4"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.42.8
-comfyui-workflow-templates==0.9.26
+comfyui-frontend-package==1.41.21
+comfyui-workflow-templates==0.9.43
 comfyui-embedded-docs==0.4.3
 torch
 torchsde