AI Rank
EN
← 开源
cswry

OSEDiff

[NeurlPS2024] One-Step Effective Diffusion Network for Real-World Image Super-Resolution

ModelsVision modelsPython
在 GitHub 打开
增长势头
+024 小时新增 Star0.0%
686
Star
42
Fork
+4
本周
1
贡献者
创建于 2024-06-03 · 更新于 2026-10-15 · 今日第 15004 名
主要开发者
cswry

cswry/OSEDiff 排名第 15,004，有 686 个星标。

同类别比较

cswry/OSEDiff 排名第 15,004，此处有 8 个同类条目；google-deepmind/recurrentgemma 排名第 15,008。

同类别比较
名称排名指标
Giphy/celeb-detection-oss14,925
语言:
Python
星标数:
692
分叉数:
68
简介:
GIPHY's Open-Source Celebrity Detection Deep Learning Model
PRBonn/LiDAR-MOS14,926
语言:
Python
星标数:
692
分叉数:
112
简介:
(LMNet) Moving Object Segmentation in 3D LiDAR Data: A Learning-based Approach Exploiting Sequential Data (RAL/IROS 2021)
IEIT-Yuan/Yuan-2.014,992
语言:
Python
星标数:
687
分叉数:
83
简介:
Yuan 2.0 Large Language Model
Osilly/Vision-DeepResearch14,994
语言:
Python
星标数:
687
分叉数:
56
简介:
[ICML 2026 & EMNLP 2026] Multimodal deep-research MLLM and benchmark. The first long-horizon multimodal deep-research MLLM, extending the number of reasoning turns to dozens and the number of search-engine interactions to hundreds.
google-deepmind/recurrentgemma15,008
语言:
Python
星标数:
686
分叉数:
39
简介:
Open weights language model from Google DeepMind, based on Griffin.
yuanxiaosc/BERT_Paper_Chinese_Translation15,030
星标数:
685
分叉数:
143
简介:
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding 论文的中文翻译 Chinese Translation!
NVlabs/OmniVinci15,032
语言:
Python
星标数:
684
分叉数:
56
简介:
OmniVinci is an omni-modal LLM for joint understanding of vision, audio, and language.
OpenMOSS/MOSS-Audio15,045
语言:
Python
星标数:
683
分叉数:
45
简介:
An open-source model for understanding speech, environmental sounds, and music through captioning, question answering, and reasoning
关键参数与规格
关键参数与规格
参数值
语言Python
类别Models
子类别Vision models
关键词Computer vision, Image production, Model architecture
星标数686
分叉数42
贡献者1
创建日期2024-06-03
更新日期2026-10-15
1 天星标数0
7 天星标数4
排名15,004
上次排名6,620
排名变化-8,384
7 天星标排名6,412
上次 7 天星标排名6,402
7 天星标排名变化-10
总星标排名15,113
上次总星标排名15,081
总星标排名变化-32
README 摘录

04)** Paper | Code 🚩 Accepted by **CVPR2026** A vision-only native generative framework for image restoration. 07)** Paper | Code 🚩 Accepted by **ICCV2025** We re-train the VAE of SD2. 1 while maintaining its UNet for more precise Real-ISR. 12)** Paper | Code 🚩 Accepted by **CVPR2025** Adjustable diffusion network for super-resolution to meet your preference! 11)** Paper | Code 🚩 Accepted by **CVPR2025** Adversarial Diffusion Compression framework for super-resolution, **offering up to 9. 09)** Paper | Code 🚩 Accepted by **TIP2026** Flexible and efficient

diffusion sampling for super-resolution, **supporting arbitrary number of inference steps**, *e. 💥💥💥Congratulations, OSEDiff has been applied to the OPPO Find X8 series! :star: If OSEDiff is helpful to your images or projects, please help star this repo. :hugs: 🎬 Overview 🔧 Dependencies and Installation 1. Download Models Dependent Models SD21 Base RAM DAPE ⚡ Quick Inference For Real-World Image Super-Resolution: For Face Restoration: 🚀 Calculate Inference Time For easy comparison, we provide the inference time testing script for osediff. If you

run the script on an A100 GPU, you would get 📏 Benchmark Results For convenient comparison, we provide OSEDiff's test results on the RealSR benchmark and DrealSR benchmark in the and directory. For the DIV2K validation benchmark, due to the large number of images, we have uploaded the test results to Google Drive. These benchmarks were directly copied from StableSR. Additionally, we also provide a script for testing IQA (Image Quality Assessment). If you run these two scripts, you would get

them at last line, which are consistent with the data presented in Table 1 of our paper. Train for Real-World Image Super-Resolution Step1: Prepare training data We use LSDIR (84991) and the first 10k images from FFHQ as our training data. Please write their absolute paths into a txt file. Each line in the txt file should be the absolute path of an image. Just like: Step2: Training for OSEDiff Pleasr put your txt file path at . Train

for Face Restoration Step1: Prepare training data We use FFHQ as our training data. Please follow the instructions aboved to prepare the txt file. Step2: Training for OSEDiff-face Pleasr put your txt file path at . If you have 4 GPUs, you can run 📷 Results For convenient evaluation and comparison, we have published the test results of DIV2K_val, RealSR, and DRealSR on Google Drive. Quantitative Comparisons (click to expand) Visual Comparisons (click to expand) 🎫 License This project is released

under the Apache 2. 📧 Contact If you have any questions, please feel free to contact: 🎓Citations statistics

常见问题

记录了多少星标和分叉？

cswry/OSEDiff 有 686 个星标和 42 个分叉。

仓库排名是多少？

cswry/OSEDiff 排名第 15,004。

记录的语言和类别是什么？

cswry/OSEDiff 有 686 个星标，语言为 Python，类别为 Models。

哪个仓库排名最接近？

cswry/OSEDiff 与 google-deepmind/recurrentgemma 比较，后者排名第 15,008，有 686 个星标。

README



One-Step Effective Diffusion Network for Real-World Image Super-Resolution

[Rongyuan Wu1,2,*](https://scholar.google.com/citations?user=A-U8zE8AAAAJ&hl=zh-CN) 
[Lingchen Sun1,2,*](https://scholar.google.com/citations?user=ZCDjTn8AAAAJ&hl=zh-CN) 
[Zhiyuan Ma1,*](https://scholar.google.com/citations?user=F15mLDYAAAAJ&hl=en) 
[Lei Zhang1,2,† ](https://www4.comp.polyu.edu.hk/~cslzhang/)




1The Hong Kong Polytechnic University, 2OPPO Research Institute 

[paper]

🚩Accepted by NeurIPS2024

😊 You may also want to check our new updates:

  • VOSR (2026.04) Paper | Code 🚩 Accepted by CVPR2026

    A vision-only native generative framework for image restoration.

  • TVT (2025.07) Paper | Code 🚩 Accepted by ICCV2025

    We re-train the VAE of SD2.1 while maintaining its UNet for more precise Real-ISR.

  • PiSA-SR (2024.12) Paper | Code 🚩 Accepted by CVPR2025

    Adjustable diffusion network for super-resolution to meet your preference!

  • AdcSR (2024.11) Paper | Code 🚩 Accepted by CVPR2025

    Adversarial Diffusion Compression framework for super-resolution, offering up to 9.3× speedup over OSEDiff.

  • CCSRv2 (2024.09) Paper | Code 🚩 Accepted by TIP2026

    Flexible and efficient diffusion sampling for super-resolution, supporting arbitrary number of inference steps, e.g., 1~15 step(s).

🔥 News

  • [2024.12] Updated OSEDiff-SD21base-face.
  • [2024.10] Updated training codes and paper. 💥💥💥Congratulations, OSEDiff has been applied to the OPPO Find X8 series!
  • [2024.07] Release OSEDiff-SD21base.
  • [2024.06] This repo is created.

:star: If OSEDiff is helpful to your images or projects, please help star this repo. Thanks! :hugs:

🎬 Overview

🔧 Dependencies and Installation

  1. Clone repo

    git clone https://github.com/cswry/OSEDiff.git
cd OSEDiff

  2. Install dependent packages

    conda create -n OSEDiff python=3.10 -y
conda activate OSEDiff
pip install --upgrade pip
pip install -r requirements.txt

  3. Download Models

Dependent Models

⚡ Quick Inference

For Real-World Image Super-Resolution:

python test_osediff.py \
-i preset/datasets/test_dataset/input \
-o preset/datasets/test_dataset/output \
--osediff_path preset/models/osediff.pkl \
--pretrained_model_name_or_path SD21BASE_PATH \
--ram_ft_path DAPE_PATH \
--ram_path RAM_PATH

For Face Restoration:

python test_osediff.py \
-i preset/datasets/test_dataset/input_face \
-o preset/datasets/test_dataset/output_face \
--upscale 1 \
--osediff_path preset/models/osediff_face.pkl \
--pretrained_model_name_or_path SD21BASE_PATH \
--ram_ft_path DAPE_PATH \
--ram_path RAM_PATH \
--align_method nofix

🚀 Calculate Inference Time

For easy comparison, we provide the inference time testing script for osediff.

python test_inference_time.py \
--osediff_path preset/models/osediff.pkl \
--inference_iterations 500 \
--warmup_iterations 5 \
--process_size 512 \
--pretrained_model_name_or_path SD21BASE_PATH \
--ram_ft_path DAPE_PATH  \
--ram_path RAM_PATH

If you run the script on an A100 GPU, you would get

Running 5 warm-up iterations...
Starting inference for 500 iterations...
Inference: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 500/500 [00:52<00:00,  9.55it/s]
Average inference time per iteration: 0.1046 seconds.

📏 Benchmark Results

For convenient comparison, we provide OSEDiff's test results on the RealSR benchmark and DrealSR benchmark in the preset/datasets/benchmark_realsr/results_osediff and preset/datasets/benchmark_drealsr/results_osediff directory. For the DIV2K validation benchmark, due to the large number of images, we have uploaded the test results to Google Drive. These benchmarks were directly copied from StableSR. Additionally, we also provide a script for testing IQA (Image Quality Assessment).

python test_metrics.py \
--inp_imgs preset/datasets/benchmark_realsr/results_osediff \
--gt_imgs preset/datasets/benchmark_realsr/test_HR \
--log preset/datasets/benchmark_realsr/results_osediff/metrics

python test_metrics.py \
--inp_imgs preset/datasets/benchmark_drealsr/results_osediff \
--gt_imgs preset/datasets/benchmark_drealsr/test_HR \
--log preset/datasets/benchmark_drealsr/results_osediff/metrics

If you run these two scripts, you would get them at last line, which are consistent with the data presented in Table 1 of our paper.

RealSR benchmark:
===== Average Metrics for [results_osediff] =====
PSNR: 25.1511; SSIM: 0.7341; LPIPS: 0.2921; DISTS: 0.2128; CLIPIQA: 0.6693; NIQE: 5.6476; MUSIQ: 69.0896; MANIQA: 0.6331 | FID: 123.493817 | FID Runtime: 6.79 sec

DrealSR benchmark:
===== Average Metrics for [results_osediff] =====
PSNR: 27.9243; SSIM: 0.7835; LPIPS: 0.2968; DISTS: 0.2165; CLIPIQA: 0.6963; NIQE: 6.4902; MUSIQ: 64.6537; MANIQA: 0.5895 | FID: 135.296586 | FID Runtime: 8.10 sec

1. Train for Real-World Image Super-Resolution

Step1: Prepare training data

We use LSDIR (84991) and the first 10k images from FFHQ as our training data. Please write their absolute paths into a txt file. Each line in the txt file should be the absolute path of an image. Just like:

LSDIR/HR_merge/0000001.png
LSDIR/HR_merge/0000002.png
...

Step2: Training for OSEDiff

Pleasr put your txt file path at YOUR TXT FILE PATH. If you have 4 GPUs, you can run

CUDA_VISIBLE_DEVICES="0,1,2,3," accelerate launch train_osediff.py \
    --pretrained_model_name_or_path=SD21BASE_PATH \
    --ram_path=RAM_PATH \
    --learning_rate=5e-5 \
    --train_batch_size=4 \
    --gradient_accumulation_steps=1 \
    --enable_xformers_memory_efficient_attention --checkpointing_steps 500 \
    --mixed_precision='fp16' \
    --report_to "tensorboard" \
    --seed 123 \
    --output_dir=experience/osediff \
    --dataset_txt_paths_list 'YOUR TXT FILE PATH','YOUR TXT FILE PATH' \
    --dataset_prob_paths_list 1,1 \
    --neg_prompt="painting, oil painting, illustration, drawing, art, sketch, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth" \
    --cfg_vsd=7.5 \
    --lora_rank=4 \
    --lambda_lpips=2 \
    --lambda_l2=1 \
    --lambda_vsd=1 \
    --lambda_vsd_lora=1 \
    --deg_file_path="params_realesrgan.yml" \
    --tracker_project_name "train_osediff"

2. Train for Face Restoration

Step1: Prepare training data

We use FFHQ as our training data. Please follow the instructions aboved to prepare the txt file.

Step2: Training for OSEDiff-face

Pleasr put your txt file path at YOUR TXT FILE PATH. If you have 4 GPUs, you can run

CUDA_VISIBLE_DEVICES="0,1,2,3," accelerate launch train_osediff_face.py \
    --pretrained_model_name_or_path=SD21BASE_PATH \
    --ram_path=RAM_PATH \
    --learning_rate=5e-5 \
    --train_batch_size=4 \
    --gradient_accumulation_steps=1 \
    --enable_xformers_memory_efficient_attention --checkpointing_steps 500 \
    --mixed_precision='fp16' \
    --report_to "tensorboard" \
    --seed 123 \
    --output_dir=experience/osediff_face \
    --dataset_txt_paths_list 'YOUR TXT FILE PATH','YOUR TXT FILE PATH' \
    --dataset_prob_paths_list 1,1 \
    --neg_prompt="blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth" \
    --cfg_vsd=7.5 \
    --lora_rank=4 \
    --lambda_lpips=2 \
    --lambda_l2=1 \
    --lambda_vsd=1 \
    --lambda_vsd_lora=1 \
    --deg_file_path="params_codeformer.yml" \
    --tracker_project_name "train_osediff_face"

📷 Results

  • For convenient evaluation and comparison, we have published the test results of DIV2K_val, RealSR, and DRealSR on Google Drive.

Quantitative Comparisons (click to expand)

Visual Comparisons (click to expand)

🎫 License

This project is released under the Apache 2.0 license.

📧 Contact

If you have any questions, please feel free to contact: rong-yuan.wu@connect.polyu.hk

🎓Citations

@article{wu2024one,
  title={One-Step Effective Diffusion Network for Real-World Image Super-Resolution},
  author={Wu, Rongyuan and Sun, Lingchen and Ma, Zhiyuan and Zhang, Lei},
  journal={arXiv preprint arXiv:2406.08177},
  year={2024}
}

statistics

排名相邻
同类项目