h-Edit is a finetuning-free diffusion-based editing method that frames editing as a reverse-time bridge modeling problem. It leverages Doob’s h-Transform for bridge construction and Langevin Monte Carlo sampling for generating edited samples.
✅ Theoretical Guarantee - Provides both explicit and implicit forms with unique features. Math doesn't lie! 📏
🚀 Training-Free, Simple, General - Smarter edits, zero headaches! 🧠
🏆 Strong Performance, SOTA on PieBench - Tackles tough cases like a champ! 💪
🛠️ Flexible - Supports conditional scores, external reward models; the first to handle both simultaneously! 🎛️
🎯 Compatible - Works with deterministic/random inversion, P2P, MasaCtrl and Plug-n-Play or even without attention control! 🔄
🔌 Plug-and-Play - Just add a pretrained diffusion model, whether for images, text, audio, or graphs, and you're all set! ✨
We showcase h-Edit's capabilities in three settings: text-guided editing (conditional scores), face swapping (external reward models), and combined text-guided & style editing (both). Each experiment is linked below.
📢 📢 📢 If h-Edit helps your work, we’d love your feedback! Please cite our paper and giving us a ⭐ - it means a lot! 🚀
Important
If this repository is useful for your work, please cite it:
@article{nguyen2025hedit,
title={h-Edit: Effective and Flexible Diffusion-Based Editing via Doob's h-Transform},
author={Nguyen, Toan and Do, Kien and Kieu, Duc and Nguyen, Thin},
journal={arXiv preprint arXiv:2503.02187},
year={2025}
}
- Release code for Face Swapping
- Release code for Combined Text-Guided & Style Editing
- Develop Webpage
- Deploy on Hugging Face
- Build App Demo