🎨 Image-To-Anime

Convert photos and videos into anime-style art with AnimeGANv3 (Hayao style) running on ONNX Runtime.

Two small scripts wrap an AnimeGANv3 generator exported to ONNX: imagetoanime.py stylises a single image, and videotoanime.py stylises a whole video frame by frame. The Hayao-style model (AnimeGANv3_Hayao_36.onnx) is bundled in the repository, inference runs on CPU by default, and CUDA GPU inference is available via a flag.

✨ Features

• Photo → anime conversion via OpenCV + ONNX Runtime, output saved at the input's original resolution
• Video → anime conversion: frame-by-frame stylisation re-encoded to MP4 (mp4v) at the source FPS and resolution
• Bundled model — AnimeGANv3_Hayao_36.onnx ships in the repo; any other AnimeGANv3 .onnx model can be supplied with -m (models with "tiny" in the filename automatically switch to 16-pixel alignment)
• CPU by default, GPU optional — -d gpu uses the CUDAExecutionProvider when a GPU-enabled onnxruntime build is detected, otherwise falls back to CPU
• Model-friendly resizing — inputs are resized to multiples of 8 (minimum 256 px) before inference
• Threaded video pipeline — a daemon reader thread decodes frames into a bounded queue (max 100) while the main loop runs inference; videos larger than 1280 px on the long edge are downscaled for inference and upscaled back on output to protect GPU memory

📋 Requirements

• Python 3
• Third-party packages (no requirements.txt; derived from imports) — install:

pip install opencv-python numpy onnxruntime Pillow tqdm

• Pillow ≥ 9.1 (videotoanime.py uses Image.Resampling)
• For GPU inference: an onnxruntime-gpu build with CUDA

🛠 Installation

git clone https://github.com/mkamranr/Image-To-Anime.git
cd Image-To-Anime
pip install opencv-python numpy onnxruntime Pillow tqdm

The model AnimeGANv3_Hayao_36.onnx is included in the repository root — no separate download is needed. Note that both scripts' default -m value points to a models/ subfolder, so either create it and move the model there:

mkdir models && mv AnimeGANv3_Hayao_36.onnx models/

or pass the model path explicitly with -m AnimeGANv3_Hayao_36.onnx.

🚀 Usage

Image

python imagetoanime.py -i photo.jpg -o anime.jpg -m AnimeGANv3_Hayao_36.onnx -d cpu

Argument	Default	Description
-i, --input_img_path	/home/ada/test_data (placeholder — always pass your own)	input image
-m, --model_path	models/AnimeGANv3_Hayao_36.onnx	ONNX model file
-o, --output_image_path	outputimage.jpg	output image
-d, --device	cpu	cpu or gpu

Prints the image size and processing time, then writes the stylised image at the original resolution.

Video

python videotoanime.py -i input.mp4 -o out/anime.mp4 -m AnimeGANv3_Hayao_36.onnx -d cpu

Argument	Default	Description
-i, --input_video_path	vid\1.mp4 (placeholder — always pass your own)	input video
-m, --model_path	models\AnimeGANv3_Hayao_36.onnx	ONNX model file
-o, --output_file_path	dist\outputvideo.mp4	output MP4 (its directory must already exist)
-d, --device	cpu	cpu or gpu

Prints a countdown of remaining frames while running, then output video: <path> when done.

⚙️ How it works

Each frame or image is resized so both dimensions are divisible by 8 (16 for "tiny" model variants, minimum 256 px), converted BGR → RGB and normalised to [-1, 1]. The AnimeGANv3 generator is executed through an onnxruntime.InferenceSession (CUDA provider when -d gpu and a GPU build is available, otherwise CPU). The network output is denormalised back to 8-bit RGB and resized to the source resolution. For video, a daemon thread reads and preprocesses frames into a bounded queue while the main loop runs inference and writes stylised frames to a cv2.VideoWriter (mp4v codec) at the source FPS; if the stream breaks mid-way the script reports "The video is broken, please upload the video again."

📁 Project structure

File	Purpose
imagetoanime.py	Single-image conversion CLI
videotoanime.py	Video conversion CLI (threaded frame pipeline)
AnimeGANv3_Hayao_36.onnx	Bundled AnimeGANv3 generator, Hayao style
LICENSE	GPL-3.0

⚠️ Notes / Limitations

• Only the Hayao-style model is bundled; other AnimeGANv3 models must be obtained separately and passed via -m.
• The default input paths are development leftovers (/home/ada/test_data, vid\1.mp4) — always supply -i.
• The output directory for videos is not created automatically, and the default output paths use Windows-style backslashes.
• -d gpu silently falls back to CPU if onnxruntime was not built with GPU support; videotoanime.py pins CUDA_VISIBLE_DEVICES=0.
• Exceptions are caught and printed rather than raised, so the scripts exit with status 0 even on failure.

📄 License

This project is licensed under the GNU General Public License v3.0 — see LICENSE.