Fanghua-Yu/SUPIR

Built on a two-stage diffusion architecture, SUPIR integrates LLaVA vision-language models for semantic understanding with SDXL's diffusion backbone and dual CLIP encoders to guide restoration. The framework offers dual model variants (Q for quality, F for fidelity) with configurable EDM sampling, classifier-free guidance, and wavelet-based color correction, enabling fine-grained control over quality-fidelity tradeoffs through adjustable hyperparameters like s_stage2 and s_cfg.

5,476 stars. No commits in the last 6 months.