SUPIR

Scaling model size is the central thesis of SUPIR, the CVPR 2024 restoration system from the XPixel group that pushes SDXL-based image restoration far enough to rival commercial upscalers like Magnific and Topaz. Blurry, compressed, noisy, or damaged photos come back sharp and richly detailed, with restoration guided by text: LLaVA can caption the degraded input automatically, and users can steer results with their own description of what the content should be. Two checkpoints serve different tastes, SUPIR-v0Q for strong generalization and generative quality and SUPIR-v0F for lighter degradations with higher fidelity. The official pipeline is research code: a conda environment, manual downloads of the SDXL base model, LLaVA, and SUPIR checkpoints, and path configuration in CKPT_PTH.py, with the README citing about 12 GB of VRAM for diffusion using half precision and tiling flags plus 16 GB to run LLaVA, so many users reach it through the popular ComfyUI-SUPIR wrapper instead. The code ships under a strict non-commercial declaration, and commercial use requires written permission from the authors.