Upscale and restore old, low-resolution family photos to high-resolution sharp images.
Denoise and sharpen blurry or noisy photos for professional-quality output.
Batch-process a folder of damaged images through an AI restoration pipeline via command line or browser UI.
Requires a multi-GPU machine with at least 12GB VRAM and several multi-gigabyte model files downloaded before first use.
SUPIR is a research project from a group of academic institutions in China and Australia, published at the CVPR 2024 computer vision conference. It is a system for restoring damaged or low-quality photographs, meaning it can take a blurry, noisy, or low-resolution image and produce a sharper, cleaner, more detailed version of it. The underlying approach combines two technologies. One is a large image-generation model called Stable Diffusion XL, which provides a strong prior about what realistic images look like. The other is a visual understanding model called LLaVA, which analyzes the image and generates a text description that guides the restoration process. By using the text description as context, the system can produce restorations that are more consistent with the actual content of the original photo rather than inventing arbitrary details. Setting up SUPIR requires downloading several large model files, some of which are multiple gigabytes in size. The setup instructions assume familiarity with Python environments and command-line tools. Running the restoration process requires a machine with multiple graphics cards due to the memory demands of the models involved. The README describes a lower-memory configuration for machines with 12 to 16 gigabytes of GPU memory, but this mode is slower. There is also an online version of the tool at a website called SupPixel AI for people who want to try it without running the software locally. The project offers two pre-trained models. One is optimized for general image quality across a range of restoration challenges. The other is better suited for photos with light degradation, where the goal is to preserve as much original detail as possible rather than re-generate it from scratch. Users run restoration jobs by pointing the tool at a folder of input images and specifying an output folder, an upscaling factor, and various quality control parameters. A browser-based interface using a library called Gradio is also available as an alternative to the command-line workflow.
← fanghua-yu on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.