PixGamma

The problem

Images that need enlarging are everywhere: product shots for an online store, old photos from the family album, wildlife and landscapes headed for a large print, illustrations, screenshots for documentation. They are often small, soft or compressed several times over. Popular AI upscalers tend to fail in familiar ways: text gets redrawn as other characters, skin, fur and fabric get covered in invented texture, colors drift, and the photos have to be uploaded to someone else's server.

The solution

PixGamma is a macOS app written in Swift that runs entirely through Core ML on the Apple Silicon GPU: photos never leave the Mac. A Light and Dark interface lets the photo fill the window: a small dock for modes, a filmstrip below, before/after at real pixels at 100% and 200% (split or side by side), and export for one photo or all. It has an Auto mode that analyses each image to pick the model and strength, a pixgamma command for batch work and a local HTTP API for scripts, websites and other apps.

What it is for

Our own photo model

PixGamma Photo v2 is the default photo model: Photo v1 trained further with a light GAN for crisper fur, hair and fabric grain (DISTS on sharp photos 0.170 vs 0.183 for v1). Text regions still run on the GAN-free Photo v1, so the GAN never draws letter strokes.

PixGamma Photo v1 is a DAT2 model fine-tuned for 15,000 iterations on the Mac itself, starting from the strongest of 10 public models we benchmarked. Training data is CC0 and public domain, low-resolution inputs follow a web-like downscale + double JPEG chain, and there is no GAN so it does not invent texture. On 40 photos never used for training: PSNR 30.5 dB (base 29.0), DISTS 0.165 (0.173), almost no color drift.

Text protection, Vietnamese included

Apple Vision finds text on the source, a 2x enlargement and the upscaled result. Each text region is measured from the pixels and rebuilt with a faithful model. Text too small or too noisy to rebuild correctly keeps the plain upscale instead of being redrawn as other letters. On the Vietnamese and English banner benchmark: 91.1% OCR accuracy.

Diffusion-based Recover

Realistic Vision 5.1 + ControlNet Tile on Core ML, seamless tiling, a hand-written DPM++ 2M SDE sampler and a dedicated eye pass. Apple Vision describes the photo on the Mac (for example "parrot, detailed feathers") for the prompt; animals, plants and portraits get Resemblance 85 so fur and hair follow the photo. Photos with text, maps or products get a safety lock that caps creativity and keeps shapes.

Shipping and updates

DMG installer and auto-update from GitHub Releases with SHA-256, bundle id and version checks and a rollback-safe app swap. A model only ships when it beats the current one on both the benchmark and real photos.

Current limits

• A 4x upscale from a very small image (under ~500 px) with dense foliage or ornament can look painterly; for print, a 2x upscale from a larger source looks more natural.
• Text that is too small (about 3 px tall in the input) cannot be rebuilt: PixGamma keeps a soft upscale instead of guessing letters. Text Apple Vision cannot find gets no dedicated protection.
• Face restoration keeps black-and-white and sepia photos uncoloured.
• Photo v2 makes fur and hair clearer than v1, but at 4x from a small image it is not yet strand-for-strand like the original: detail fully gone from the input cannot be recreated; sharpness is still being improved model by model.

PixGamma is free for non-commercial use and runs entirely on the Mac. Sample images on this page come from Wikimedia Commons (CC0 and public domain).