Hybrid Regularization Improves Diffusion-based Inverse Problem Solving (ICLR 2025) Paper

This is the README file for the implementation of Hybrid Regularization Improves Diffusion-based Inverse Problem Solving (HRDIS).

🌟Comparison between RED-diff and our proposed HRDIS.

✨HRDIS can generate diverse and clear reconstructions.

Installation

Download ImageNet and FFHQ dataset. You need to write your data directory at _configs\dataset\imagenet256_val(ffhq256_val).yaml.

name: "ImageNet_256x256"
root: "./data/imagenet"
split: "val"
image_size: 256
channels: 3
meta_root: "_exp"
transform: "diffusion"
subset_txt: "misc/dgp_top1k.txt"  

Download pretrained checkpoints and put them in _exp/ckpts as the following file.

Dataset	File	Model Source
ImageNet	imagenet/256x256_diffusion_uncond.pt	guided-diffusion
FFHQ	ffhq/ffhq_10m.pt	DPS

Install the dependencies:

pip install -r requirements.txt

Git clone external codes for non-linear deblurring.

git clone https://github.com/VinAIResearch/blur-kernel-space-exploring bkse

Download mask for inapinting from the Palette and put it in _exp/masks/20ff.npz

Inference

Select the test dataset by enabling the following code in ./main.py

@hydra.main(version_base="1.2", config_path="_configs", config_name="imagenet256_uncond")
#@hydra.main(version_base="1.2", config_path="_configs", config_name="ffhq256_uncond")

Tune the hyperparameters interactively using the sampling script:

sh sample_test.sh 

Reference

@inproceedings{
dou2025hybrid,
title={Hybrid Regularization Improves Diffusion-based Inverse Problem Solving},
author={Hongkun Dou and Zeyu Li and Jinyang Du and Lijun Yang and Wen Yao and Yue Deng},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=d7pr2doXn3}
}