SUPIR

Overview

SUPIR isn’t just another photo fixer. It’s like hitting “undo” on photo damage but with AI brains doing the heavy lifting. Instead of sliders and brushes you just type what the photo should look like. Like “elderly man smiling with sunglasses” and the thing gets to work.

A crew from big research spots in China built it up. We’re talking Shenzhen Institute of Advanced Tech Tencent ARC and a few major universities. These folks already did solid stuff like ESRGAN so this isn’t their first AI rodeo.

How it works under the hood

SUPIR’s built on Stable Diffusion XL which is a big 2.6B parameter model. They chopped ControlNet down and added a custom plug-in called ZeroSFT. Then trained a strong encoder that can handle nasty real-world damage in pics.

The model got fed 20 million captioned images. That way it “learned” to connect what things look like with how they’re described in words. So when you type “toddler in a red sweater with big smile” it knows what to aim for.

Here’s what sets it apart.

Text prompt control. You tell it what to aim for in words.

Negative prompts. You can block stuff like “messy” or “blurry” from showing up.

Smart sampling. It doesn’t just make something pretty it tries to stick close to the original damaged pic.

Understands both words and images. So it doesn’t get confused by weird damage or weird lighting.

Can handle real-world mess. Old family pics water-damaged prints low-res junk from your 2007 phone camera… bring it.

Good resolution. Goes up to 1024x1024 which is solid.

SUPIR's Licence

The SUPIR project on GitHub runs on a custom open license that lets you use copy change and share the software but only if it’s not for money. That rule makes it a non-commercial open license which sets it apart from regular ones like MIT Apache 2.0 or GPL that do allow making money off it.