SUPIR

We'd reach for SUPIR when a photo has lost texture, not just pixels. A 1994 family scan that has gone soft and smeared is exactly the case where diffusion reconstruction pulls plausible skin, fabric and foliage back out of nowhere, and where a text prompt like "natural skin, sharp eyes" actually changes the result. Prompt-guided control is the reason to pick it over a generic upscaler. Pick the self-hosted GitHub model if you have a capable GPU, want to iterate on the prompt, or need to run restorations offline. Pick SupPixel AI at suppixel.ai if you want the same model without the environment setup — XPixel Group launched it officially as the hosted front door to SUPIR. Pass on it if your workflow is high-volume batch processing, needs deterministic output that repeats identically run to run, or lives inside an automated pipeline. Diffusion restoration is a per-image craft, not a conveyor belt. Video and real-time footage are out of scope too. The closest alternative depends on what you're optimising for. Topaz Gigapixel and hosted restoration APIs win on throughput, repeatability and vendor support commitments. SUPIR wins on reconstruction quality and on the fact that it's open source — you can inspect it, fine-tune it, and run it without a subscription. One caveat worth stating plainly: because it generates detail rather than recovering it, SUPIR can invent texture that wasn't in the original. For forensic, evidentiary or archival-fidelity work where every pixel must trace back to the source, that's a disqualifier. For creative restoration, it's the feature. In practice we'd keep both paths installed: SupPixel AI for a quick pass on a single image, the local model when a prompt needs twenty iterations to land.