DiffBIR

Open source

What the maker publishes

Additional components

The project says it is based on ControlNet and BasicSR and uses Stable Diffusion v2.1 for training stage two. source

Apple Silicon

The project announced Apple Silicon support and refers users to instructions for CPU, CUDA, and MPS devices. source

Batch processing

Yes source

Deployment

The instructions clone the repository, create a Python 3.10 Conda environment, install requirements, and run inference locally. source

Founded

2023 source

Headquarters

China source

License

The project is released under the Apache 2.0 license. source

Maximum scale

8x source

Model downloads

Pretrained weights are linked from Hugging Face, Baidu Netdisk, and OpenXLab, and the README says inference weights download automatically. source

Model integrations

Inference uses BSRNet for super-resolution, SwinIR-Face for face restoration, and SCUNet-PSNR for denoising as stage-one models. source

Online demo

The README announces an online demo on OpenXLab for the general and face models. source

Purpose

DiffBIR is a blind image restoration pipeline that uses a generative diffusion prior. source

Quality trade-off

The README says using the condition as the sampling start point can improve stability but may decrease sample quality. source

Restoration tasks

The repository provides inference examples for blind image super-resolution, aligned and unaligned face restoration, and blind image denoising. source

Support

The README directs questions to [email protected]. source

Tiled sampling

Tiled sampling supports super-resolution at large scale factors on low-VRAM graphics cards. source

Training limitation

The README states that the stage-two training script does not support a validation set. source

Upscaling method

ai source

Versions

The README lists DiffBIR v2 and v2.1, and says v2.1 includes a model trained on Unsplash captions generated with LLaVA, more samplers, and improved tiled-sampling support. source

Checked against the maker's pages on 12 Oct 2026.