---
base_model:
- Wan-AI/Wan2.1-I2V-14B-480P
language:
- en
library_name: diffusers
license: mit
pipeline_tag: image-to-video
---

<div align="center">
<img src='https://github.com/TencentARC/ToonComposer/raw/main/samples/ToonComposer-Icon.png' width='120px'>
</div>

<p align="center"> <b> ToonComposer: Streamlining Cartoon Production with Generative Post-Keyframing </b> </p>

<p align="center"> <a href='https://lg-li.github.io/project/tooncomposer'><img src='https://img.shields.io/badge/Project-Page-Green'></a> &nbsp;
 <a href='https://huggingface.co/TencentARC/ToonComposer'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-blue'></a> &nbsp;
 <a href="https://arxiv.org/abs/2508.10881"><img src="https://img.shields.io/static/v1?label=ArxivPreprint&message=ToonComposer&color=red&logo=arxiv"></a> &nbsp;
 <a href="https://github.com/TencentARC/ToonComposer"><img src="https://img.shields.io/static/v1?label=Github&message=ToonComposer&color=black&logo=github"></a>
</p>

<div align="center">
<img src='https://github.com/TencentARC/ToonComposer/raw/main/samples/ToonComposer-TLDR.jpg' width='550px'>
</div>

Traditional cartoon/anime production is time-consuming, requiring skilled artists for keyframing, inbetweening, and colorization.
ToonComposer streamlines this with generative AI, turning hours of manual work of inbetweening and colorization into a single, seamless process. Visit our [project page](https://lg-li.github.io/project/tooncomposer) and read our [paper](https://arxiv.org/abs/2508.10881) for more details.
This HF model repo provides model weights of ToonComposer. Codes are available at [our GitHub repo](https://github.com/TencentARC/ToonComposer).

## Sample Usage

You can use the `ToonComposer` model with the `diffusers` library. Ensure your environment has the required dependencies, including `flash-attn` as specified in the [official GitHub repository](https://github.com/TencentARC/ToonComposer) for optimal performance.

```python
import torch
from diffusers import DiffusionPipeline
from PIL import Image
import numpy as np # For creating dummy images if paths are not found

# Load the ToonComposer pipeline from the Hugging Face Hub
pipeline = DiffusionPipeline.from_pretrained(
    "TencentARC/ToonComposer",
    torch_dtype=torch.float16, # Use torch.bfloat16 for newer GPUs or if preferred
    trust_remote_code=True      # Required to load custom pipeline code
)
pipeline.to("cuda") # Move model to GPU for faster inference (can use "cpu" for CPU inference)

# --- Prepare your input data ---
# 1. Initial Colored Keyframe (Reference Image)
# This image sets the base visual style and initial frame.
# Replace 'path/to/your/initial_colored_keyframe.png' with your actual image file.
try:
    initial_colored_keyframe = Image.open("path/to/your/initial_colored_keyframe.png").convert("RGB")
except FileNotFoundError:
    print("Warning: Initial colored keyframe image not found. Using a dummy white image for demonstration.")
    # Create a dummy white image matching target resolution (e.g., 1088x608 from model config)
    initial_colored_keyframe = Image.fromarray(np.full((608, 1088, 3), 255, dtype=np.uint8))

# 2. Sketch Keyframe (for motion control)
# This is typically a black and white line drawing that guides motion at a specific frame.
# Replace 'path/to/your/sketch_at_frame_X.png' with your actual sketch image path.
# ToonComposer supports multiple sketches at different time steps. For this example, we use one.
try:
    sketch_keyframe = Image.open("path/to/your/sketch_at_frame_X.png").convert("RGB")
except FileNotFoundError:
    print("Warning: Sketch keyframe image not found. Using a dummy black image for demonstration.")
    # Create a dummy black image matching target resolution
    sketch_keyframe = Image.fromarray(np.full((608, 1088, 3), 0, dtype=np.uint8))

# Text Prompt: Describe the desired motion or scene
prompt = "a joyful character bouncing a ball"

# Video Generation Parameters (adjust as needed)
# Refer to the model's config.json or official GitHub for recommended values.
num_frames = 33      # Example number of frames from model's config.json
height = 608         # Example resolution from model's config.json
width = 1088         # Example resolution from model's config.json
guidance_scale = 7.5 # Common value for text-to-image/video diffusion

# --- Generate the video frames ---
# The exact arguments for this custom pipeline might vary.
# We infer a plausible API based on common diffusion model practices and the paper's description.
# The `image` argument is for the initial colored keyframe, and `sketches` for the list of sketch images.
video_frames = pipeline(
    prompt=prompt,
    image=initial_colored_keyframe,
    sketches=[(5, sketch_keyframe)], # Example: Apply `sketch_keyframe` at frame index 5
    num_frames=num_frames,
    height=height,
    width=width,
    guidance_scale=guidance_scale,
).frames # The output is expected to be a list of PIL Images.

# --- Save or display the generated video frames ---
output_dir = "./tooncomposer_output"
import os
os.makedirs(output_dir, exist_ok=True)
for i, frame in enumerate(video_frames):
    frame.save(f"{output_dir}/frame_{i:04d}.png")
print(f"Generated {len(video_frames)} frames to '{output_dir}'.")

# Optional: To compile frames into a GIF (requires `imageio` and `imageio-ffmpeg` to be installed)
# import imageio
# try:
#     imageio.mimsave(f"{output_dir}/output_video.gif", video_frames, fps=10)
#     print(f"Saved GIF to '{output_dir}/output_video.gif'.")
# except Exception as e:
#     print(f"Could not save GIF (ensure imageio and imageio-ffmpeg are installed): {e}")
```

## Citation

If you find ToonComposer useful, please consider citing:

```
@article{li2025tooncomposer,
  title={ToonComposer: Streamlining Cartoon Production with Generative Post-Keyframing},
  author={Li, Lingen and Wang, Guangzhi and Zhang, Zhaoyang and Li, Yaowei and Li, Xiaoyu and Dou, Qi and Gu, Jinwei and Xue, Tianfan and Shan, Ying},
  journal={arXiv preprint arXiv:2508.10881},
  year={2025}
}
```