Changelog | PixGamma docs

Notable changes by version, newest first. The full technical log is the repository's CHANGELOG.md.

Unreleased

New interface, Light and Dark: the photo fills the window, a small dock holds only mode, scale and Xem trước (preview); a filmstrip sits below and the tuning column on the right collapses. Fit shows the whole photo with the preview area outlined; 100% and 200% compare before/after at real pixels, split or side by side. Two export buttons: this photo, or all of them. Follows the system theme, or pick Light / Dark from the View menu.
PixGamma Photo v2 is the default photo model: Photo v1 trained further with a light GAN. DISTS 0.164 vs 0.165 on the benchmark, 0.170 vs 0.183 on sharp photos. Text regions keep Photo v1, so the GAN never draws letter strokes.
Recover understands the scene: Apple Vision describes the photo (for example "parrot, detailed feathers") for the prompt; before, feathers could come out as fabric. Resemblance defaults to 85 for animals, plants and portraits and 40 otherwise; the app has a switch to set it by hand.
Safety lock for Recover: signs, maps and text-heavy photos are capped at creativity 0.3 with the structure lock on; product photos at 0.45. Text is protected as whole paragraphs.

Face restoration coloured black-and-white photos: a sepia 1910 portrait got blue eyes and pink skin. Monochrome photos now keep their own tone in the face.

Pixora is now PixGamma (the Pixora name is used by several other AI photo apps), with a new lens-and-aperture logo. The app is PixGamma.app, the command is pixgamma, and the photo model is PixGamma Photo v1.
Old installs keep updating: releases also carry Pixora-<version>.dmg for 0.2.1 apps, and the new updater renames Pixora.app to PixGamma.app. Application Support/Pixora moves to PixGamma; the pixora command and PIXORA_* variables keep working.

Soft, wobbly keycap letters (for example Y): text regions now run through PixGamma Photo v1 instead of the old text model, and the small-text fade is gone. OCR 89.2% → 91.1%, PSNR 29.2 → 30.5 dB, sharpness 0.58 → 0.69; +1.2 dB on a real keyboard photo.
Slow text pass: 2 flip/rotation passes instead of 8. A 480 px keyboard to 4x went from 37 s to 12 s and reads better.
Recover redrew keycap letters (A S D F G H J became A S D Y G N I): Recover now finds text like the other modes; color keyboard photos keep every key.
Recover killed by macOS on large images: each tile's Core ML outputs are freed right away. 1440 px needs 8.5 GB instead of 28.6 GB; 3436 × 5194 finishes with 10.2 GB.

Every mode, 1x–16x, --width/--height, every slider, gpu/ane/all/cpu, on 1 px, 48 px, 2000 × 60 strips, 3000 × 2000, grayscale, CMYK, transparent, palette, TIFF, WebP and HEIC inputs.

DMG installer with the CLI and every upscaling model; Recover models in 2 zip files.
Auto-update from GitHub Releases: daily check, DMG download, SHA-256, bundle id and version checks, app swap with rollback and relaunch. Read-only GitHub token kept in the Keychain.

PixGamma Photo v1, our own photo model (DAT2 fine-tuned from 4xBHI_dat2_real): PSNR 30.5 dB vs 29.0 dB, DISTS 0.165 vs 0.173, no color drift.
Recover mode: Realistic Vision 5.1 + ControlNet Tile on Core ML, seamless MultiDiffusion, a dedicated eye pass, regional prompts, Creativity and Resemblance sliders.
Vietnamese text recognition (vi-VT); text read is written to the report.
Dedicated text model (RealPLKSR) trained in three rounds on synthetic Vietnamese and English text.
Dehalo around high-contrast edges.
Synthetic text dataset, per-line and per-size text benchmark.

Auto picked Art & CGI for white-background product photos (22/22 Lenovo photos now use the photo model).
Black laptop details looked lighter and hazy after upscaling: Auto no longer sharpens or adds contrast or grain on its own.
Soft, haloed letters after text protection; broken blocks at the end of small text lines; the CLI could not find models when called from PATH; replacing a model with the same name kept the old compiled one.

Three-pane SwiftUI app: drag-and-drop image list, preview with a before/after slider and 3 zoom levels, adjustments panel.
Five Core ML modes: Auto, Standard, High Fidelity, Art & CGI, Low Resolution.
Realistic (CodeFormer) and Creative (GFPGAN) face restoration.
Color preservation, text protection with Apple Vision, context-aware tiling.
pixgamma CLI and local HTTP API.