Breaking

Retro Diffusion Developer Releases pixel-bench, an Open-Source Benchmark for AI Pixel Art Reconstruction

July 16, 2026 at 22:47 EDT

Open Source
Media Generation
Research & Papers

A developer behind Retro Diffusion has published pixel-bench, an open-source, MIT-licensed benchmark that objectively compares tools designed to turn distorted, AI-generated "pixel art" back into true native pixel imagery. The project, hosted on GitHub, lets anyone add their own reconstruction methods, run reproducible tests, and measure results independently.

Open-Source Benchmark · Retro Diffusion

A Standard Yardstick for Restoring Fake Pixel Art

pixel-bench is a new MIT-licensed benchmark that objectively measures which tools best turn distorted, AI-generated "pixel art" back into true, grid-aligned native imagery — reproducibly, and open for anyone to extend.

distortion categories in the v1 suite

built-in reconstruction methods to compare

MIT

open-source license, community-extendable

One image folder becomes a test battery

100 supplied ground-truth images × 43 deterministic distortions = 4,300 reproducible test inputs.

100

images

×43 →

4,300

test inputs

How it runs

pip install -e .

Set up the benchmark; optional backends auto-skipped if missing.

→

pixelbench run

Run all methods against your folder of true 1x pixel art.

→

pixelbench report

Generate a Markdown report, per-category and overall.

5 built-in methods

fixer — Retro Diffusion Pixel Art Fixer (non-ML)
snapper — PixelSnapper port
pixeldetector — PixelDetector port
unfake
naive — baseline

Add new methods in a single file with a @register decorator.

What it measures

Resolution — exact, ±1px, ±5%, error
Color fidelity — delta E (CIELAB), PSNR
Placement — exact-match pixels, grid alignment
Palette — color-count error, palette delta E

Higher is better — except relative error and delta E, where lower wins.

Worth watching: the vendor's own Pixel Art Fixer is among the benchmarked methods — but the open, deterministic, version-controlled design exists precisely so anyone can verify results independently. Its value now hinges on community adoption and shared submissions.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.