REG
[ICML'25] REG: Rectified Gradient Guidance for Conditional Diffusion Models
Install / Use
/learn @zhengqigao/REGREADME
REG
[ICML'25] REG: Rectified Gradient Guidance for Conditional Diffusion Models
TL;DR: Previous studies have identified a discrepancy between guidance theory and its practical implementation. In this work, we propose a refined explanation for guidance theory, inspired by the observation that the original theory focuses on scaling marginal distributions, whereas the correct formulation should aim to scale the joint distribution.
Instruction
REG requires only minimal modifications—just a one-line change in the reverse denoising process—so it should be straightforward to implement.
- For the ImageNet experiments in the paper, we forked DiT and EDMv2 and applied our modifications.
- For the text-to-image experiments, we modify the source code of Hugging Face's
diffuserslibrary.
Please refer to each subfolder for detailed instructions. Note that there may be minor typos in the code, as the files are recreated based on my local codebase and are intentionally kept minimal for clarity. If you encounter any issues, feel free to open an issue or contact me at zhengqi@mit.edu.
