Freedom251 › Image Upscaling Software
DiffBIR
by DiffBIR
What DiffBIR is
DiffBIR is free, open-source image restoration software using a generative diffusion prior. Its examples cover blind image super-resolution, aligned and unaligned face restoration, and blind image denoising. The project lists upscaling up to 8x and batch processing. Tiled sampling is intended for large scale factors on low-VRAM graphics cards. The repository lists versions 2 and 2.1; version 2.1 adds a model trained on Unsplash captions generated with LLaVA, additional samplers, and improved tiled-sampling support. Stage-one inference uses BSRNet for super-resolution, SwinIR-Face for face restoration, and SCUNet-PSNR for denoising. The project is based on ControlNet and BasicSR and uses Stable Diffusion v2.1 for stage-two training. Local setup involves cloning the repository, creating a Python 3.10 Conda environment, installing requirements, and running inference. Apple Silicon support is announced for CPU, CUDA, and MPS devices. It is released under Apache 2.0.
Who it suits
DiffBIR suits users who can run open-source restoration software locally on Linux, macOS, self-hosted, or web environments. It is relevant for super-resolution, face restoration, and denoising workflows.
What it does well
- Supports super-resolution, face restoration, and denoising
- Upscaling listed up to 8x
- Batch processing is supported
- Tiled sampling targets low-VRAM graphics cards
- Released under Apache 2.0
What to know first
- GPU memory can limit inference
- Local setup requires Python 3.10 Conda
- Stage-two training lacks validation-set support
- Condition-start sampling may reduce sample quality
Plans
| Plan | Billed and limits | Price |
|---|---|---|
| DiffBIR | Open-source research code and pretrained weights; GPU memory can limit inference | Free |
As github.com lists them, read 12 Oct 2026. The maker's own page has the final word.
Verdict
DiffBIR provides several restoration tasks, multiple model components, tiled sampling, and announced Apple Silicon support. Running it requires a local environment, and available GPU memory can constrain inference.
Facts from DiffBIR's own pages
- Additional components
- The project says it is based on ControlNet and BasicSR and uses Stable Diffusion v2.1 for training stage two. · github.com, 12 Oct 2026
- Apple Silicon
- The project announced Apple Silicon support and refers users to instructions for CPU, CUDA, and MPS devices. · github.com, 12 Oct 2026
- Batch processing
- Yes · github.com, 22 Sept 2026
- Deployment
- The instructions clone the repository, create a Python 3.10 Conda environment, install requirements, and run inference locally. · github.com, 12 Oct 2026
- Founded
- 2023 · github.com, 28 Sept 2026
- Headquarters
- China · github.com, 28 Sept 2026
- License
- The project is released under the Apache 2.0 license. · github.com, 12 Oct 2026
- Maximum scale
- 8x · github.com, 22 Sept 2026
- Model downloads
- Pretrained weights are linked from Hugging Face, Baidu Netdisk, and OpenXLab, and the README says inference weights download automatically. · github.com, 12 Oct 2026
- Model integrations
- Inference uses BSRNet for super-resolution, SwinIR-Face for face restoration, and SCUNet-PSNR for denoising as stage-one models. · github.com, 12 Oct 2026
- Online demo
- The README announces an online demo on OpenXLab for the general and face models. · github.com, 12 Oct 2026
- Purpose
- DiffBIR is a blind image restoration pipeline that uses a generative diffusion prior. · github.com, 12 Oct 2026
- Quality trade-off
- The README says using the condition as the sampling start point can improve stability but may decrease sample quality. · github.com, 12 Oct 2026
- Restoration tasks
- The repository provides inference examples for blind image super-resolution, aligned and unaligned face restoration, and blind image denoising. · github.com, 12 Oct 2026
- Support
- The README directs questions to [email protected]. · github.com, 12 Oct 2026
- Tiled sampling
- Tiled sampling supports super-resolution at large scale factors on low-VRAM graphics cards. · github.com, 12 Oct 2026
- Training limitation
- The README states that the stage-two training script does not support a validation set. · github.com, 12 Oct 2026
- Upscaling method
- ai · github.com, 22 Sept 2026
- 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. · github.com, 12 Oct 2026
Where DiffBIR stands on Freedom251
- Image Upscaling Software#27 of 99