📢 News

• 2026/08/02: Demo Code is released.
• 2026/06/30: Our paper is available at arXiv.

👀 Overview

We warp the input image to enlarge small salient regions (e.g., objects, faces, eyes) to better preserve fine details in the compressed latent space during image-to-image translation. In the figure above, original latents are shown in 🟥, and warped latents in 🟩.

🌟 Highlights

• Detail-preserving — preserves fine details in latent diffusion models.
• Model-agnostic — no architectural modifications needed.
• Ultra-efficient — only 0.006s additional latency and zero extra learnable parameters.

🎬 Videos

Please check our Project Page for more details.

🤗 HuggingFace Demo

Try the interactive demo directly in your browser — no installation needed:

👉 https://huggingface.co/spaces/ShenZheng2000/WarpI2I-demo

📦 Model: https://huggingface.co/ShenZheng2000/WarpI2I

The demo covers all three tasks via separate tabs:

Tab	Input	Options
Human Relighting	Portrait photo	Moonlight / Golden Sunlight / Foggy / Noon Sunlight
Driving Relighting	Driving scene	Foggy / Golden Sunlight
Driving Domain Transfer	Driving scene	Day→Night / Clear→Rainy / Cityscapes→Foggy / Cityscapes→Dark

To run the demo locally:

pip install gradio huggingface_hub
python app.py
# then open http://127.0.0.1:7860

⚙️ Installation

See docs/install.md.

📥 Dataset Preparation

See docs/dataset.md.

🚀 Training and Testing

See docs/train_test.md.

🔌 Warp–Unwarp Integration

Key warp–unwarp code insertions are marked with ✅.

The main insertion points are:

• Data loading → src/my_utils/training_utils.py (PairedDataset, UnpairedDataset)
• Pix2Pix-Turbo → src/train_pix2pix_turbo.py
• CycleGAN-Turbo → src/train_cyclegan_turbo.py
• Warp utilities → src/warp_utils/

✍️ Citation

If you find this work useful, please cite:

@inproceedings{zheng2026warpi2i,
  title     = {WarpI2I: Image Warping for Image-to-Image Translation},
  author    = {Zheng, Shen and Ghosh, Anurag and Parmar, Gaurav and Narasimhan, Srinivasa},
  booktitle = {ECCV},
  year      = {2026}
}

🤝 Contact

If you have any questions, feel free to raise an issue (recommended) or send an email to [email protected].

🙏 Acknowledgement

This project builds upon the following excellent open-source works:

• img2img-turbo
• Instance-Warp
• Two-Plane Prior