Algolzw/FoD has 32 GitHub stars and 2 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-21 19:50 UTC. Last push 2026-10-02.
Progression mesurée des étoiles
- in 24h
- 0
- in 7 days
- 0
- in 15 days
- Historique en cours
- in 1 month
- Historique en cours
Variations nettes sur des périodes qui se chevauchent, pas des totaux quotidiens. Les mesures absentes restent vides. Les étoiles indiquent l’intérêt, pas la qualité.
À propos
Efficient Image Restoration with State-Dependent Forward Diffusion, TMLR 2026
Apprentissage automatiqueREADME.md
OriginalEfficient Image Restoration with State-Dependent Forward Diffusion (FoD)
Official PyTorch Implementation
How to Run the Code?
Dependencies
- OS: Ubuntu 20.04
- nvidia:
- cuda: 12.2
- python 3.11
Install
We advise you first create a virtual environment with:
python3 -m venv .env
source .env/bin/activate
pip install -U pip
pip install -r requirements.txt
Image Restoration
Here we provide an example for image deraining task, but can be changed to any problem by replacing the dataset.
Dataset Preparation
Download dataset and process it in a way such that rain images and no-rain images are in separately directories, as
#### for training dataset ####
datasets/rain/trainH/GT
datasets/rain/trainH/LQ
#### for testing dataset ####
datasets/rain/testH/GT
datasets/rain/testH/LQ
Then get into the image_restoration directory and modify the dataset paths in train_IR.py and sample_IR.py. You can also change the run-name and task-name to save models in different directories.
Train
The core algorithms for FoD is in image_restoration/diffusion/fod_diffusion.py.
You can train the model for image restoration following below bash scripts:
cd image_restoration
# For single GPU:
torchrun --nnodes=1 --nproc_per_node=1 --master_port=34567 train_IR.py --global-batch-size 16
# Change the nproc_per_node and global-batch-size for multi-GPU training
torchrun --nnodes=1 --nproc_per_node=4 --master_port=34567 train_IR.py --global-batch-size 64
Then the models and training logs will save in results/{task-name}.
Evaluation
To evaluate our method, please modify the benchmark path and model path and run
python sample_IR.py
Unconditional Generation
Train: Similar to the Image Restoration task, you can train the model on CIFAR-10 following below bash script:
cd image_generation
# For single GPU:
torchrun --nnodes=1 --nproc_per_node=1 --master_port=34567 train_cifar10.py --global-batch-size 128
# Change the nproc_per_node and global-batch-size for multi-GPU training
torchrun --nnodes=1 --nproc_per_node=4 --master_port=34567 train_cifar10.py --global-batch-size 128
Sampling: You need to first change the pretrained model path and sample images with the sample_cifar10.py:
python sample_cifar10.py
Evaluation
We include a ddp_sample_cifar10.py script which samples a large number of images from a pretrain FoD model in parallel. Run:
torchrun --nnodes=1 --nproc_per_node=1 --master_port=45678 ddp_sample_cifar10.py
To calculate FID, Inception Score, etc., you need to obtain a cifar_train.npz of cifar10, and run:
python scripts/evaluator.py cifar_train.npz samples/cifar10/U-Net-0500000-seed-0.npz
Results
FoD forward diffusion process (click to expand)

Image Restoration Results (click to expand)

Acknowledgment: Our FoD is based on DiT and guided-diffusion . Thanks for their code!
Contact
If you have any question, please contact: [email protected]
Citations
If our code helps your research or work, please consider citing our paper. The following are BibTeX references:
@article{luo2025forward,
title={Forward-only Diffusion Probabilistic Models},
author={Luo, Ziwei and Gustafsson, Fredrik K and Sj{\"o}lund, Jens and Sch{\"o}n, Thomas B},
journal={arXiv preprint arXiv:2505.16733},
year={2025}
}
--- Thanks for your interest! ---
statistics
Explorer des projets similaires
Chargement…
