SUPIR

SUPIR is a diffusion-based image restoration model from XPixel Group that rebuilds low-quality photos, portraits, game stills and vintage images at high fidelity, using text prompts to guide the reconstruction.

SUPIR is an image restoration model from XPixel Group built on a large-scale diffusion generative prior. Rather than applying a fixed sharpening or denoising filter, it reconstructs a damaged or low-quality image by drawing on what a diffusion model has learned about how real photographs look, then uses text prompts to steer the result toward the content the user expects. The project frames its output as high-fidelity general restoration: the goal is a plausible, detailed image rather than a mechanically cleaned one. The showcase material organizes restoration by subject, and the range hints at where the model is strongest. Landscape shots, portraits and animal photos get dedicated treatment, with the portrait section emphasising fine facial detail and the animal section emphasising fur texture. Separate sections cover gaming screenshots and stills from older films, plus vintage family photographs. That spread suggests the intended users are archivists, photo retouchers, game preservationists and film restorers who work with degraded or low-resolution source material and want a single model rather than a per-scene filter chain. In practice the tool is text-driven. A user supplies a low-quality input and describes what the restored version should contain, and the model fills in structure and texture consistent with both the source and the prompt. This is the main point of difference from conventional upscalers, which cannot be told what they are looking at. The trade-off is that prompt quality matters: the description shapes the output, so restoration becomes partly an editorial act rather than a purely automatic one. SUPIR began as a research release, and the team has since announced SupPixel AI, a hosted product built on the same technology for image processing and upscaling. Buyers choosing between SUPIR and mainstream enhancement tools should weigh its generative approach: it can recover detail that filter-based upscalers cannot invent, but that same invention means results are interpretations of the original, not faithful recoveries. It suits users who want vivid, detailed restoration and are comfortable with a model that makes judgement calls.