SUPIR

SUPIR restores and enlarges photos with diffusion models. It's for photographers and image researchers who want to process degraded images on their own hardware and control how closely the results follow the originals.

• Restoration and upscaling: Improve degraded photos and increase their resolution, with models suited to general restoration or lighter degradation where preserving detail matters more.
• Fidelity controls: Adjust the balance between visual appeal and faithfulness to the input, since settings that produce more attractive results can also change original details.
• Prompt guidance: Use positive and negative text prompts to influence the image's appearance, alongside controls for color correction.
• Batch and interactive use: Process folders of images or use a Gradio browser interface with an image comparison slider.

The Python software runs locally with downloaded model weights. It uses SDXL and LLaVA, and supports Juggernaut-XL as an alternative SDXL base. Its GPU workflow uses CUDA; the reduced-memory mode lists 12 GB of VRAM for diffusion and 16 GB for LLaVA, with slower processing. SupPixel AI is a separate hosted app built on SUPIR for online image processing and upscaling.

SUPIR restricts use, reproduction and distribution to non-commercial purposes. Commercial use requires prior written permission from Dr. Jinjin Gu.