init commit

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text

app.py

@@ -0,0 +1,260 @@
+import time
+import gradio as gr
+import torch as th
+import torch
+import numpy as np
+import tempfile
+from diffusers import AutoencoderKLWan
+from diffusers.utils import export_to_video, load_image
+from diffusers.schedulers import UniPCMultistepScheduler
+from transformers import CLIPVisionModel
+from chronoedit_diffusers.pipeline_chronoedit import ChronoEditPipeline
+from chronoedit_diffusers.transformer_chronoedit import ChronoEditTransformer3DModel
+from PIL import Image
+from huggingface_hub import hf_hub_download
+from prompt_enhancer import enhance_prompt
+from transformers import (
+    Qwen2_5_VLForConditionalGeneration, 
+    AutoProcessor, 
+)
+
+
+th.enable_grad(False)
+th.backends.cuda.preferred_linalg_library(backend="magma")
+
+
+start = time.time()
+
+model_id = "nvidia/ChronoEdit-14B-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(
+	model_id,
+	subfolder="image_encoder",
+	torch_dtype=torch.float32
+)
+print("✓ Loaded image encoder")
+
+vae = AutoencoderKLWan.from_pretrained(
+	model_id,
+	subfolder="vae",
+	torch_dtype=torch.bfloat16
+)
+print("✓ Loaded VAE")
+
+transformer = ChronoEditTransformer3DModel.from_pretrained(
+	model_id,
+	subfolder="transformer",
+	torch_dtype=torch.bfloat16
+)
+print("✓ Loaded transformer")
+
+pipe = ChronoEditPipeline.from_pretrained(
+	model_id,
+	image_encoder=image_encoder,
+	transformer=transformer,
+	vae=vae,
+	torch_dtype=torch.bfloat16
+)
+print("✓ Created pipeline")
+
+lora_path = hf_hub_download(repo_id=model_id, filename="lora/chronoedit_distill_lora.safetensors")
+# Load LoRA if specified
+if lora_path:
+	print(f"Loading LoRA weights from {lora_path}...")
+	pipe.load_lora_weights(lora_path)
+
+	pipe.fuse_lora(lora_scale=1.0)
+	print(f"✓ Fused LoRA with scale 1.0")
+# Setup scheduler
+pipe.scheduler = UniPCMultistepScheduler.from_config(
+	pipe.scheduler.config,
+	flow_shift=2.0
+)
+print(f"✓ Configured scheduler (flow_shift=2.0)")
+
+# Move to device
+pipe.to("cuda:0")
+print(f"✓ Models loaded and moved to cuda:0")
+
+end = time.time()
+print(f"Model loaded in {end - start:.2f}s.")
+
+start = time.time()
+prompt_enhancer_model = "Qwen/Qwen2.5-VL-7B-Instruct"
+prompt_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+	prompt_enhancer_model,
+	torch_dtype=torch.bfloat16,
+	attn_implementation="flash_attention_2",
+	device_map="cuda:1",
+)
+processor = AutoProcessor.from_pretrained(prompt_enhancer_model)
+end = time.time()
+print(f"Prompt enhancer loaded in {end - start:.2f}s.")
+
+
+def calculate_dimensions(image,  mod_value):
+    """
+    Calculate output dimensions based on resolution settings.
+    
+    Args:
+        image: PIL Image
+        mod_value: Modulo value for dimension alignment
+        
+    Returns:
+        Tuple of (width, height)
+    """
+    
+    # Get max area from preset or override 
+    target_area = 720 * 1280
+    
+    # Calculate dimensions maintaining aspect ratio
+    aspect_ratio = image.height / image.width
+    calculated_height = round(np.sqrt(target_area * aspect_ratio)) // mod_value * mod_value
+    calculated_width = round(np.sqrt(target_area / aspect_ratio)) // mod_value * mod_value
+    
+    return calculated_width, calculated_height
+
+
+def run_inference(
+	image_path: str,
+	prompt: str,
+	enable_temporal_reasoning: bool,
+	num_inference_steps: int = 8,
+	guidance_scale: float = 1.0,
+	shift: float = 2.0,
+	num_temporal_reasoning_steps: int = 8,
+):
+	pipe.to("cuda:0")
+
+	# Rewriter (optional)
+	final_prompt = prompt
+
+	# Enhance prompt with CoT reasoning
+	cot_prompt = enhance_prompt(
+		image_path,
+		prompt,
+		prompt_model,
+		processor,
+	)
+
+	# Print enhanced CoT prompt
+	print("\n" + "=" * 80)
+	print("Enhanced CoT Prompt:")
+	print("=" * 80)
+	print(cot_prompt)
+	print("=" * 80 + "\n")
+	final_prompt = cot_prompt
+
+	# Inference
+	print(f"Loading input image: {image_path}")
+	image = load_image(image_path)
+	mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+	width, height = calculate_dimensions(
+		image,
+		mod_value
+	)
+	print(f"Output dimensions: {width}x{height}")
+	image = image.resize((width, height))
+	num_frames = 29 if enable_temporal_reasoning else 5
+	with th.no_grad():
+		start = time.time()
+		output = pipe(
+			image=image,
+			prompt=final_prompt,
+			height=height,
+			width=width,
+			num_frames=num_frames,
+			num_inference_steps=num_inference_steps,
+			guidance_scale=guidance_scale,
+			enable_temporal_reasoning=enable_temporal_reasoning,
+			num_temporal_reasoning_steps=num_temporal_reasoning_steps,
+			offload_model=True,
+		).frames[0]
+
+		end = time.time()
+
+	video_tmp = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
+	output_path_video = video_tmp.name
+	video_tmp.close()
+	image_tmp = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
+	output_path_image = image_tmp.name
+	image_tmp.close()
+	export_to_video(output, output_path_video, fps=10)
+	Image.fromarray((output[-1] * 255).clip(0, 255).astype("uint8")).save(output_path_image)
+
+	log_text = (
+		f"Final prompt: {final_prompt}\n"
+		f"Guidance: {guidance_scale}, Shift: {shift}, Steps: {num_inference_steps}\n"
+		f"Inference: {end - start:.2f}s"
+	)
+	if enable_temporal_reasoning:
+		log_text += f"Temporal reasoning: {enable_temporal_reasoning}, Steps: {num_temporal_reasoning_steps}\n"
+	return output_path_image, output_path_video	#, log_text
+
+
+def build_ui() -> gr.Blocks:
+	with gr.Blocks(title="ChronoEdit", theme=gr.themes.Soft()) as demo:
+
+		gr.Markdown("""
+		# 🚀 ChronoEdit Demo		
+		[[Project Page]](https://research.nvidia.com/labs/toronto-ai/chronoedit/) | 
+		[[Code]](https://github.com/nv-tlabs/ChronoEdit) |
+		[[Technical Report]](https://arxiv.org/abs/2510.04290)
+		""")
+		
+		with gr.Row():
+			image = gr.Image(type="filepath", label="Input Image")
+			output_image = gr.Image(label="Generated Image")
+		with gr.Row():
+			with gr.Column(scale=1):
+				prompt = gr.Textbox(label="Prompt", lines=4, value="")
+				enable_temporal_reasoning = gr.Checkbox(label="Enable temporal reasoning", value=False)
+				run_btn = gr.Button("Start Generation", variant="primary")
+			with gr.Column(scale=1):
+				output_video = gr.Video(label="Temporal Reasoning Visualization", visible=False)
+		
+		# with gr.Row():
+		# 	num_inference_steps = gr.Slider(minimum=4, maximum=75, step=1, value=50, label="Num Inference Steps")
+		# 	guidance_scale = gr.Slider(minimum=1.0, maximum=10.0, step=0.5, value=1.0, label="Guidance Scale")
+		# 	shift = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=5.0, label="Shift")
+		# 	num_temporal_reasoning_steps = gr.Slider(minimum=0, maximum=50, step=1, value=50, label="Number of temporal reasoning steps")
+
+		# log_text = gr.Markdown("Logs will appear here.")	
+
+		def _on_run(image_path, prompt, enable_temporal_reasoning):
+			image_out_path, video_out_path = run_inference(
+				image_path=image_path,
+				prompt=prompt,
+				enable_temporal_reasoning=enable_temporal_reasoning,
+			)
+			video_update = gr.update(visible=enable_temporal_reasoning, value=(video_out_path if enable_temporal_reasoning else None))
+			return image_out_path, video_update
+
+		run_btn.click(
+			_on_run,
+			inputs=[image, prompt, enable_temporal_reasoning],
+			outputs=[output_image, output_video] #, log_text],
+		)
+
+		gr.Examples(
+			examples=[ 
+				[
+					"examples/1.png",
+					"The user wants to change the scene so that the girl in the traditional-style painting, wearing her ornate floral robe and headdress, is now playing a guitar. Her graceful appearance remains unchanged\u2014smooth black hair tied neatly, soft facial features with a calm, focused expression\u2014but her pose shifts: both hands are engaged with the guitar. One hand rests on the neck of the instrument, fingers pressing the strings with delicate precision, while the other hand strums near the sound hole. The guitar is positioned naturally across her lap, blending with the elegance of her posture. The traditional painting style is preserved, but the addition of the guitar introduces a modern contrast, giving the scene a harmonious fusion of classical refinement and contemporary music.",
+					False,
+				],
+				[
+					"examples/1.png",
+					"The user wants to change the scene so that the girl in the traditional-style painting, wearing her ornate floral robe and headdress, is now playing a guitar. Her graceful appearance remains unchanged\u2014smooth black hair tied neatly, soft facial features with a calm, focused expression\u2014but her pose shifts: both hands are engaged with the guitar. One hand rests on the neck of the instrument, fingers pressing the strings with delicate precision, while the other hand strums near the sound hole. The guitar is positioned naturally across her lap, blending with the elegance of her posture. The traditional painting style is preserved, but the addition of the guitar introduces a modern contrast, giving the scene a harmonious fusion of classical refinement and contemporary music.",
+					True,
+				],
+			],
+			inputs=[image, prompt, enable_temporal_reasoning], outputs=[output_image, output_video], fn=_on_run, cache_examples=True
+		)
+
+	return demo
+
+
+# CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=$(pwd) python scripts/app.py
+if __name__ == "__main__":
+	demo = build_ui()
+	demo.launch(server_name="0.0.0.0", server_port=7869)

chronoedit_diffusers/pipeline_chronoedit.py

@@ -0,0 +1,764 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import PIL
+import regex as re
+import torch
+from transformers import AutoTokenizer, CLIPImageProcessor, CLIPVisionModel, UMT5EncoderModel
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.image_processor import PipelineImageInput
+from diffusers.loaders import WanLoraLoaderMixin
+from diffusers.models import AutoencoderKLWan, WanTransformer3DModel
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.wan.pipeline_output import WanPipelineOutput
+from chronoedit_diffusers.transformer_chronoedit import ChronoEditTransformer3DModel
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+if is_ftfy_available():
+    import ftfy
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> import numpy as np
+        >>> from diffusers import AutoencoderKLWan, WanImageToVideoPipeline
+        >>> from diffusers.utils import export_to_video, load_image
+        >>> from transformers import CLIPVisionModel
+
+        >>> # Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers
+        >>> model_id = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+        >>> image_encoder = CLIPVisionModel.from_pretrained(
+        ...     model_id, subfolder="image_encoder", torch_dtype=torch.float32
+        ... )
+        >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+        >>> pipe = WanImageToVideoPipeline.from_pretrained(
+        ...     model_id, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+        ... )
+        >>> max_area = 480 * 832
+        >>> aspect_ratio = image.height / image.width
+        >>> mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+        >>> height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+        >>> width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+        >>> image = image.resize((width, height))
+        >>> prompt = (
+        ...     "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in "
+        ...     "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+        ... )
+        >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+        >>> output = pipe(
+        ...     image=image,
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     height=height,
+        ...     width=width,
+        ...     num_frames=81,
+        ...     guidance_scale=5.0,
+        ... ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=16)
+        ```
+"""
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class ChronoEditPipeline(DiffusionPipeline, WanLoraLoaderMixin):
+    r"""
+    Pipeline for image-to-video generation using Wan.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        tokenizer ([`T5Tokenizer`]):
+            Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer),
+            specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        image_encoder ([`CLIPVisionModel`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModel), specifically
+            the
+            [clip-vit-huge-patch14](https://github.com/mlfoundations/open_clip/blob/main/docs/PRETRAINED.md#vit-h14-xlm-roberta-large)
+            variant.
+        transformer ([`WanTransformer3DModel`]):
+            Conditional Transformer to denoise the input latents.
+        scheduler ([`UniPCMultistepScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->image_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: UMT5EncoderModel,
+        image_encoder: CLIPVisionModel,
+        image_processor: CLIPImageProcessor,
+        transformer: ChronoEditTransformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            image_encoder=image_encoder,
+            transformer=transformer,
+            scheduler=scheduler,
+            image_processor=image_processor,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+        self.image_processor = image_processor
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 512,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt = [prompt_clean(u) for u in prompt]
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_image(
+        self,
+        image: PipelineImageInput,
+        device: Optional[torch.device] = None,
+    ):
+        device = device or self._execution_device
+        image = self.image_processor(images=image, return_tensors="pt").to(device)
+        image_embeds = self.image_encoder(**image, output_hidden_states=True)
+        return image_embeds.hidden_states[-2]
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        negative_prompt,
+        image,
+        height,
+        width,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        image_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if image is not None and image_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `image`: {image} and `image_embeds`: {image_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        if image is None and image_embeds is None:
+            raise ValueError(
+                "Provide either `image` or `prompt_embeds`. Cannot leave both `image` and `image_embeds` undefined."
+            )
+        if image is not None and not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image):
+            raise ValueError(f"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` but is {type(image)}")
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif negative_prompt is not None and (
+            not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list)
+        ):
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+    def prepare_latents(
+        self,
+        image: PipelineImageInput,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        latent_height = height // self.vae_scale_factor_spatial
+        latent_width = width // self.vae_scale_factor_spatial
+
+        shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device=device, dtype=dtype)
+
+        image = image.unsqueeze(2)
+        video_condition = torch.cat(
+            [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 1, height, width)], dim=2
+        )
+        video_condition = video_condition.to(device=device, dtype=dtype)
+
+        latents_mean = (
+            torch.tensor(self.vae.config.latents_mean)
+            .view(1, self.vae.config.z_dim, 1, 1, 1)
+            .to(latents.device, latents.dtype)
+        )
+        latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+            latents.device, latents.dtype
+        )
+
+        if isinstance(generator, list):
+            latent_condition = [
+                retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") for _ in generator
+            ]
+            latent_condition = torch.cat(latent_condition)
+        else:
+            latent_condition = retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax")
+            latent_condition = latent_condition.repeat(batch_size, 1, 1, 1, 1)
+
+        latent_condition = (latent_condition - latents_mean) * latents_std
+
+        mask_lat_size = torch.ones(batch_size, 1, num_frames, latent_height, latent_width)
+        mask_lat_size[:, :, list(range(1, num_frames))] = 0
+        first_frame_mask = mask_lat_size[:, :, 0:1]
+        first_frame_mask = torch.repeat_interleave(first_frame_mask, dim=2, repeats=self.vae_scale_factor_temporal)
+        mask_lat_size = torch.concat([first_frame_mask, mask_lat_size[:, :, 1:, :]], dim=2)
+        mask_lat_size = mask_lat_size.view(batch_size, -1, self.vae_scale_factor_temporal, latent_height, latent_width)
+        mask_lat_size = mask_lat_size.transpose(1, 2)
+        mask_lat_size = mask_lat_size.to(latent_condition.device)
+
+        return latents, torch.concat([mask_lat_size, latent_condition], dim=1)
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: PipelineImageInput,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        image_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+        enable_temporal_reasoning: bool = False,
+        num_temporal_reasoning_steps: int = 0,
+        offload_model: bool=False
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PipelineImageInput`):
+                The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`.
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            height (`int`, defaults to `480`):
+                The height of the generated video.
+            width (`int`, defaults to `832`):
+                The width of the generated video.
+            num_frames (`int`, defaults to `81`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `5.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `negative_prompt` input argument.
+            image_embeds (`torch.Tensor`, *optional*):
+                Pre-generated image embeddings. Can be used to easily tweak image inputs (weighting). If not provided,
+                image embeddings are generated from the `image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, *optional*, defaults to `512`):
+                The maximum sequence length of the prompt.
+            shift (`float`, *optional*, defaults to `5.0`):
+                The shift of the flow.
+            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
+                The dtype to use for the torch.amp.autocast.
+        Examples:
+
+        Returns:
+            [`~WanPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where
+                the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            image,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            image_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        if num_frames % self.vae_scale_factor_temporal != 1:
+            logger.warning(
+                f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number."
+            )
+            num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
+        num_frames = max(num_frames, 1)
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if offload_model:
+            self.text_encoder.cpu()
+        # Encode image embedding
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        if image_embeds is None:
+            image_embeds = self.encode_image(image, device)
+        image_embeds = image_embeds.repeat(batch_size, 1, 1)
+        image_embeds = image_embeds.to(transformer_dtype)
+
+        if offload_model:
+            self.image_encoder.cpu()
+
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.vae.config.z_dim
+        image = self.video_processor.preprocess(image, height=height, width=width).to(device, dtype=torch.bfloat16)
+        latents, condition = self.prepare_latents(
+            image,
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.bfloat16,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        if offload_model:
+            torch.cuda.empty_cache()
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+
+                if self.interrupt:
+                    continue
+
+                if enable_temporal_reasoning and i == num_temporal_reasoning_steps:
+                    latents = latents[:, :, [0, -1]]
+                    condition = condition[:, :, [0, -1]]
+
+                    for j in range(len(self.scheduler.model_outputs)):
+                        if self.scheduler.model_outputs[j] is not None:
+                            if latents.shape[-3] != self.scheduler.model_outputs[j].shape[-3]:
+                                self.scheduler.model_outputs[j] = self.scheduler.model_outputs[j][:,:,[0, -1]]
+                    if self.scheduler.last_sample is not None:
+                        self.scheduler.last_sample = self.scheduler.last_sample[:, :, [0, -1]]
+
+                self._current_timestep = t
+                latent_model_input = torch.cat([latents, condition], dim=1).to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_hidden_states_image=image_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if offload_model:
+                    torch.cuda.empty_cache()
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        encoder_hidden_states_image=image_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if offload_model:
+            self.transformer.cpu()
+            torch.cuda.empty_cache()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+
+            if enable_temporal_reasoning and num_temporal_reasoning_steps > 0:
+                video_edit = self.vae.decode(latents[:, :, [0, -1]], return_dict=False)[0]
+                video_reason = self.vae.decode(latents[:, :, :-1], return_dict=False)[0]
+                video = torch.cat([video_reason, video_edit[:, :, 1:]], dim=2)
+            else:
+                video = self.vae.decode(latents, return_dict=False)[0]
+
+            # video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return WanPipelineOutput(frames=video)

chronoedit_diffusers/transformer_chronoedit.py

@@ -0,0 +1,476 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
+from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import Attention
+from diffusers.models.cache_utils import CacheMixin
+from diffusers.models.embeddings import PixArtAlphaTextProjection, TimestepEmbedding, Timesteps, get_1d_rotary_pos_embed
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import FP32LayerNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class ChronoEditAttnProcessor2_0:
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("ChronoEditAttnProcessor2_0 requires PyTorch 2.0. To use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        encoder_hidden_states_img = None
+        if attn.add_k_proj is not None:
+            encoder_hidden_states_img = encoder_hidden_states[:, :257]
+            encoder_hidden_states = encoder_hidden_states[:, 257:]
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(hidden_states: torch.Tensor, freqs: torch.Tensor):
+                x_rotated = torch.view_as_complex(hidden_states.to(torch.float64).unflatten(3, (-1, 2)))
+                x_out = torch.view_as_real(x_rotated * freqs).flatten(3, 4)
+                return x_out.type_as(hidden_states)
+
+            query = apply_rotary_emb(query, rotary_emb)
+            key = apply_rotary_emb(key, rotary_emb)
+
+        # I2V task
+        hidden_states_img = None
+        if encoder_hidden_states_img is not None:
+            key_img = attn.add_k_proj(encoder_hidden_states_img)
+            key_img = attn.norm_added_k(key_img)
+            value_img = attn.add_v_proj(encoder_hidden_states_img)
+
+            key_img = key_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+            value_img = value_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+            hidden_states_img = F.scaled_dot_product_attention(
+                query, key_img, value_img, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+            hidden_states_img = hidden_states_img.transpose(1, 2).flatten(2, 3)
+            hidden_states_img = hidden_states_img.type_as(query)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        if hidden_states_img is not None:
+            hidden_states = hidden_states + hidden_states_img
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
+class ChronoEditImageEmbedding(torch.nn.Module):
+    def __init__(self, in_features: int, out_features: int):
+        super().__init__()
+
+        self.norm1 = FP32LayerNorm(in_features)
+        self.ff = FeedForward(in_features, out_features, mult=1, activation_fn="gelu")
+        self.norm2 = FP32LayerNorm(out_features)
+
+    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.norm1(encoder_hidden_states_image)
+        hidden_states = self.ff(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+        return hidden_states
+
+
+class ChronoEditTimeTextImageEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        time_freq_dim: int,
+        time_proj_dim: int,
+        text_embed_dim: int,
+        image_embed_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.timesteps_proj = Timesteps(num_channels=time_freq_dim, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.time_embedder = TimestepEmbedding(in_channels=time_freq_dim, time_embed_dim=dim)
+        self.act_fn = nn.SiLU()
+        self.time_proj = nn.Linear(dim, time_proj_dim)
+        self.text_embedder = PixArtAlphaTextProjection(text_embed_dim, dim, act_fn="gelu_tanh")
+
+        self.image_embedder = None
+        if image_embed_dim is not None:
+            self.image_embedder = ChronoEditImageEmbedding(image_embed_dim, dim)
+
+    def forward(
+        self,
+        timestep: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+    ):
+        timestep = self.timesteps_proj(timestep)
+
+        time_embedder_dtype = next(iter(self.time_embedder.parameters())).dtype
+        if timestep.dtype != time_embedder_dtype and time_embedder_dtype != torch.int8:
+            timestep = timestep.to(time_embedder_dtype)
+        temb = self.time_embedder(timestep).type_as(encoder_hidden_states)
+        timestep_proj = self.time_proj(self.act_fn(temb))
+
+        encoder_hidden_states = self.text_embedder(encoder_hidden_states)
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)
+
+        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image
+
+
+class ChronoEditRotaryPosEmbed(nn.Module):
+    def __init__(
+        self, attention_head_dim: int, patch_size: Tuple[int, int, int], max_seq_len: int, theta: float = 10000.0, temporal_skip_len: int = 8
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.patch_size = patch_size
+        self.max_seq_len = max_seq_len
+        self.temporal_skip_len = temporal_skip_len
+
+        h_dim = w_dim = 2 * (attention_head_dim // 6)
+        t_dim = attention_head_dim - h_dim - w_dim
+
+        freqs = []
+        for dim in [t_dim, h_dim, w_dim]:
+            freq = get_1d_rotary_pos_embed(
+                dim, max_seq_len, theta, use_real=False, repeat_interleave_real=False, freqs_dtype=torch.float64
+            )
+            freqs.append(freq)
+        self.freqs = torch.cat(freqs, dim=1)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w
+
+        self.freqs = self.freqs.to(hidden_states.device)
+        freqs = self.freqs.split_with_sizes(
+            [
+                self.attention_head_dim // 2 - 2 * (self.attention_head_dim // 6),
+                self.attention_head_dim // 6,
+                self.attention_head_dim // 6,
+            ],
+            dim=1,
+        )
+
+        assert num_frames == 2 or num_frames == self.temporal_skip_len, f"num_frames must be 2 or {self.temporal_skip_len}, but got {num_frames}"
+        if num_frames == 2:
+            freqs_f = freqs[0][:self.temporal_skip_len][[0, -1]].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        else:
+            freqs_f = freqs[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_h = freqs[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_w = freqs[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+        freqs = torch.cat([freqs_f, freqs_h, freqs_w], dim=-1).reshape(1, 1, ppf * pph * ppw, -1)
+        return freqs
+
+
+class ChronoEditTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        ffn_dim: int,
+        num_heads: int,
+        qk_norm: str = "rms_norm_across_heads",
+        cross_attn_norm: bool = False,
+        eps: float = 1e-6,
+        added_kv_proj_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        # 1. Self-attention
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            processor=ChronoEditAttnProcessor2_0(),
+        )
+
+        # 2. Cross-attention
+        self.attn2 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            added_kv_proj_dim=added_kv_proj_dim,
+            added_proj_bias=True,
+            processor=ChronoEditAttnProcessor2_0(),
+        )
+        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+
+        # 3. Feed-forward
+        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
+        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        rotary_emb: torch.Tensor,
+    ) -> torch.Tensor:
+        shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+            self.scale_shift_table + temb.float()
+        ).chunk(6, dim=1)
+
+        # 1. Self-attention
+        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+        attn_output = self.attn1(hidden_states=norm_hidden_states, rotary_emb=rotary_emb)
+        hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+
+        # 2. Cross-attention
+        norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+        attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+        hidden_states = hidden_states + attn_output
+
+        # 3. Feed-forward
+        norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
+            hidden_states
+        )
+        ff_output = self.ffn(norm_hidden_states)
+        hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+
+        return hidden_states
+
+
+class ChronoEditTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+    r"""
+    A Transformer model for video-like data used in the ChronoEdit model.
+
+    Args:
+        patch_size (`Tuple[int]`, defaults to `(1, 2, 2)`):
+            3D patch dimensions for video embedding (t_patch, h_patch, w_patch).
+        num_attention_heads (`int`, defaults to `40`):
+            Fixed length for text embeddings.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_dim (`int`, defaults to `512`):
+            Input dimension for text embeddings.
+        freq_dim (`int`, defaults to `256`):
+            Dimension for sinusoidal time embeddings.
+        ffn_dim (`int`, defaults to `13824`):
+            Intermediate dimension in feed-forward network.
+        num_layers (`int`, defaults to `40`):
+            The number of layers of transformer blocks to use.
+        window_size (`Tuple[int]`, defaults to `(-1, -1)`):
+            Window size for local attention (-1 indicates global attention).
+        cross_attn_norm (`bool`, defaults to `True`):
+            Enable cross-attention normalization.
+        qk_norm (`bool`, defaults to `True`):
+            Enable query/key normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        add_img_emb (`bool`, defaults to `False`):
+            Whether to use img_emb.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["patch_embedding", "condition_embedder", "norm"]
+    _no_split_modules = ["ChronoEditTransformerBlock"]
+    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
+    _keys_to_ignore_on_load_unexpected = ["norm_added_q"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: Tuple[int] = (1, 2, 2),
+        num_attention_heads: int = 40,
+        attention_head_dim: int = 128,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        text_dim: int = 4096,
+        freq_dim: int = 256,
+        ffn_dim: int = 13824,
+        num_layers: int = 40,
+        cross_attn_norm: bool = True,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        eps: float = 1e-6,
+        image_dim: Optional[int] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        rope_max_seq_len: int = 1024,
+        rope_temporal_skip_len: int = 8,
+    ) -> None:
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        # 1. Patch & position embedding
+        self.rope = ChronoEditRotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len, temporal_skip_len=rope_temporal_skip_len)
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Condition embeddings
+        # image_embedding_dim=1280 for I2V model
+        self.condition_embedder = ChronoEditTimeTextImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=freq_dim,
+            time_proj_dim=inner_dim * 6,
+            text_embed_dim=text_dim,
+            image_embed_dim=image_dim,
+        )
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList(
+            [
+                ChronoEditTransformerBlock(
+                    inner_dim, ffn_dim, num_attention_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output norm & projection
+        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image
+        )
+        timestep_proj = timestep_proj.unflatten(1, (6, -1))
+
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+
+        # 4. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb
+                )
+        else:
+            for block in self.blocks:
+                hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb)
+
+        # 5. Output norm, projection & unpatchify
+        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
+
+        # Move the shift and scale tensors to the same device as hidden_states.
+        # When using multi-GPU inference via accelerate these will be on the
+        # first device rather than the last device, which hidden_states ends up
+        # on.
+        shift = shift.to(hidden_states.device)
+        scale = scale.to(hidden_states.device)
+
+        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

prompt_enhancer.py

@@ -0,0 +1,289 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import torch
+from PIL import Image
+from transformers import (
+    Qwen2_5_VLForConditionalGeneration, 
+    AutoProcessor, 
+    Qwen3VLMoeForConditionalGeneration,
+)
+from qwen_vl_utils import process_vision_info
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="Enhance a prompt with CoT reasoning given an input image and prompt"
+    )
+    parser.add_argument(
+        "--input-image",
+        type=str,
+        default="./assets/images/input.jpg",
+        help="Path to the input image (default: ./assets/images/input.jpg)"
+    )
+    parser.add_argument(
+        "--input-prompt",
+        type=str,
+        required=True,
+        help="Input prompt to enhance with CoT reasoning"
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="Qwen/Qwen3-VL-30B-A3B-Instruct",
+        choices=[
+            "Qwen/Qwen2.5-VL-7B-Instruct",
+            "Qwen/Qwen3-VL-30B-A3B-Instruct",
+        ],
+        help="Model to use for prompt enhancement"
+    )
+    parser.add_argument(
+        "--max-resolution",
+        type=int,
+        default=1080,
+        help="Maximum resolution for the shortest edge (default: 1080)"
+    )
+    return parser.parse_args()
+
+
+def pick_attn_implementation(prefer_flash: bool = True) -> str:
+    """
+    Decide the best attn_implementation based on environment.
+
+    Returns one of: "flash_attention_2", "sdpa", "eager".
+    """
+    # Try FlashAttention v2 first (needs SM80+ and the wheel to import)
+    if prefer_flash:
+        try:
+            import flash_attn  # noqa: F401
+            if torch.cuda.is_available():
+                major, minor = torch.cuda.get_device_capability()
+                # FlashAttn requires Ampere (SM80) or newer
+                if (major, minor) >= (8, 0):
+                    return "flash_attention_2"
+        except Exception:
+            pass
+    try:
+        if torch.backends.cuda.sdp_kernel.is_available():
+            return "sdpa"
+    except Exception:
+        pass
+
+    # Fallback: eager (always works, slower)
+    return "eager"
+def load_model(model_name):
+    """Load the vision-language model and processor."""
+    print(f"Loading model: {model_name}")
+
+    attn_impl = pick_attn_implementation(prefer_flash=True)
+
+    if model_name == "Qwen/Qwen2.5-VL-7B-Instruct":
+        model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+            model_name,
+            dtype=torch.bfloat16,
+            attn_implementation=attn_impl,
+            device_map="auto",
+        )
+        processor = AutoProcessor.from_pretrained(model_name)
+    
+    elif model_name == "Qwen/Qwen3-VL-30B-A3B-Instruct":
+        model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
+            model_name, 
+            dtype=torch.bfloat16,
+            attn_implementation=attn_impl,
+            device_map="auto"
+        )
+        processor = AutoProcessor.from_pretrained(model_name)
+
+    else:
+        raise ValueError(f"Unsupported model: {model_name}")
+
+    return model, processor
+
+
+def resize_if_needed(image, max_resolution=1080):
+    """Resize image so that the shortest edge is at most max_resolution pixels."""
+    width, height = image.size
+    if min(width, height) > max_resolution:
+        scaling_factor = max_resolution / float(min(width, height))
+        new_size = (int(width * scaling_factor), int(height * scaling_factor))
+        print(f"Resizing image from {image.size} to {new_size}")
+        return image.resize(new_size, Image.LANCZOS)
+    return image
+
+
+def _run_model_inference(messages, model, processor):
+    """
+    Helper function to run model inference.
+    
+    Args:
+        messages: Chat messages for the model
+        model: The loaded VL model
+        processor: The model's processor
+        
+    Returns:
+        str: Generated text
+    """
+    if isinstance(model, Qwen2_5_VLForConditionalGeneration):
+        text = processor.apply_chat_template(
+            messages, tokenize=False, add_generation_prompt=True
+        )
+        images, videos = process_vision_info(messages)
+        inputs = processor(
+            text=[text],
+            images=images,
+            videos=videos,
+            padding=True,
+            return_tensors="pt",
+        )
+        inputs = inputs.to(model.device).to(model.dtype)
+        generated_ids = model.generate(**inputs, max_new_tokens=512)
+
+    elif isinstance(model, Qwen3VLMoeForConditionalGeneration):
+        inputs = processor.apply_chat_template(
+            messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt"
+        )
+        inputs = inputs.to(model.device).to(model.dtype)
+        generated_ids = model.generate(**inputs, max_new_tokens=512)
+
+    else:
+        raise ValueError("Unsupported model type")
+
+    # Decode the generated text
+    generated_ids_trimmed = [
+        out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
+    ]
+    output_text = processor.batch_decode(
+        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+    )
+    
+    return output_text[0]
+
+
+def enhance_prompt(input_image_path, input_prompt, model, processor, max_resolution=1080):
+    """
+    Enhance a prompt with Chain-of-Thought reasoning given an input image and prompt.
+    
+    Args:
+        input_image_path: Path to the input image
+        input_prompt: The input editing instruction prompt
+        model: The loaded VL model
+        processor: The model's processor
+        max_resolution: Maximum resolution for image resizing
+        
+    Returns:
+        str: Enhanced CoT prompt
+    """
+    # Load and resize image
+    print(f"Loading image: {input_image_path}")
+    input_image = Image.open(input_image_path).convert("RGB")
+    input_image = resize_if_needed(input_image, max_resolution)
+
+    cot_prompt = f"""You are a professional edit instruction rewriter and prompt engineer. Your task is to generate a precise, concise, and visually achievable chain-of-thought reasoning based on the user-provided instruction and the image to be edited.
+
+You have the following information:
+1. The user provides an image (the original image to be edited)
+2. question text: {input_prompt}
+
+Your task is NOT to output the final answer or the edited image. Instead, you must:
+- Generate a "thinking" or chain-of-thought process that explains how you reason about the editing task.
+- First identify the task type, then provide reasoning/analysis that leads to how the image should be edited.
+- Always describe pose and appearance in detail.
+- Match the original visual style or genre (anime, CG art, cinematic, poster). If not explicit, choose a stylistically appropriate one based on the image.
+- Incorporate motion and camera direction when relevant (e.g., walking, turning, dolly in/out, pan), implying natural human/character motion and interactions.
+- Maintain quoted phrases or titles exactly (e.g., character names, series names). Do not translate or alter the original language of text.
+
+## Task Type Handling Rules:
+
+**1. Standard Editing Tasks (e.g., Add, Delete, Replace, Action Change):**
+- For replacement tasks, specify what to replace and key visual features of the new element.
+- For text editing tasks, specify text position, color, and layout concisely.
+- If the user wants to "extract" something, this means they want to remove the background and only keep the specified object isolated. We should add "while removing the background" to the reasoning.
+- Explicitly note what must stay unchanged: appearances (hairstyle, clothing, expression, skin tone/race, age), posture, pose, visual style/genre, spatial layout, and shot composition (e.g., medium shot, close-up, side view).
+
+**2. Character Consistency Editing Tasks (e.g., Scenario Change):**
+- For tasks that place an object/character (e.g., human, robot, animal) in a completely new scenario, preserve the object's core identity (appearance, materials, key features) but adapt its pose, interaction, and context to fit naturally in the new environment.
+- Reason about how the object should interact with the new scenario (e.g., pose changes, hand positions, orientation, facial direction).
+- The background and context should transform completely to match the new scenario while maintaining visual coherence.
+- Describe both what stays the same (core appearance) and what must change (pose, interaction, setting) to make the scene look realistic and natural.
+
+The length of outputs should be **around 80 - 100 words** to fully describe the transformation. Always start with "The user wants to ..."
+
+Example Output 1 (Standard Editing Task):
+The user wants to make the knight kneel on his right knee while keeping the rest of the pose intact. 
+The knight should lower his stance so his right leg bends to the ground in a kneeling position, with the left leg bent upright to support balance. 
+The shield with the NVIDIA logo should still be held up firmly in his left hand, angled forward in a defensive posture, while the right hand continues gripping the weapon. 
+The armor reflections, proportions, and medieval style should remain consistent, emphasizing a powerful and respectful kneeling stance.
+
+Example Output 2 (Character Consistency Editing Task):
+The user wants to change the image by modifying the scene so that the woman is drinking coffee in a cozy coffee shop. 
+The elegant anime-style woman keeps her same graceful expression, long flowing dark hair adorned with golden ornaments, and detailed traditional outfit with red and gold floral patterns. 
+She is now seated at a wooden café table, holding a steaming cup of coffee near her lips with one hand, while soft sunlight filters through the window, highlighting her refined features. 
+The background transforms into a warmly lit café interior with subtle reflections, bookshelves, and gentle ambience, maintaining the delicate, painterly aesthetic.
+"""
+
+    # Create messages for CoT generation
+    messages = [
+        {
+            "role": "system",
+            "content": [
+                {"type": "text", "text": cot_prompt},
+            ],
+        },
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "image": input_image},
+            ],
+        }
+    ]
+
+    # Generate CoT reasoning
+    print("Generating Chain-of-Thought enhanced prompt...")
+    cot_prompt_output = _run_model_inference(messages, model, processor)
+    
+    return cot_prompt_output
+
+
+def main():
+    args = parse_args()
+    
+    # Load model
+    model, processor = load_model(args.model)
+    
+    # Enhance prompt with CoT reasoning
+    cot_prompt = enhance_prompt(
+        args.input_image,
+        args.input_prompt,
+        model,
+        processor,
+        args.max_resolution
+    )
+    
+    # Print enhanced CoT prompt
+    print("\n" + "="*80)
+    print("Enhanced CoT Prompt:")
+    print("="*80)
+    print(cot_prompt)
+    print("="*80 + "\n")
+
+
+if __name__ == "__main__":
+    main()

requirements.txt

@@ -0,0 +1,26 @@
+torch==2.7.1
+torchvision==0.22.1 
+
+einops==0.8.1
+typing-extensions==4.14.1
+
+diffusers==0.35.2
+peft==0.17.1
+accelerate==1.8.1
+transformers==4.57.1
+sentencepiece==0.2.0
+tokenizers==0.22
+xfuser==0.4.4
+regex==2024.11.6
+ftfy==6.3.1
+numpy==1.26.4
+pillow==11.1.0
+qwen-vl-utils==0.0.14
+gradio==5.49.1
+
+# Progress bars
+tqdm==4.67.1
+
+# Video export utilities (if using diffusers export_to_video)
+imageio==2.37.0
+imageio-ffmpeg==0.6.0