PixText: a sub-1B model that fixes mangled text in AI-generated images
Creepy_Astronomer_83 · reddit · 2026-10-23
Built for the frustration of images that look great except for a few mangled words or Chinese characters, PixText refines text directly in pixel space (avoiding VAE compression loss on small text), is under 1B parameters (0.95B), needs no text encoder or VAE, and uses MeanFlow to repair text in just 4 sampling steps.
In ComfyUI you can draw boxes and type the correct text, or start with automatic word detection and adjust the suggested boxes; a comparison node shows each region before and after. Available via the ComfyUI Registry, ready-made workflows, GitHub code, and Hugging Face weights (ATH-MaaS/PixText).
More from Multimodal
- Small open-source TTS is getting interesting: Hojo-TTS-Light ships 40M and 80M variants — ZabihullahAtal · 2026-10-23
- Obsidian Light Mirage prompt template sculpts dramatic chiaroscuro portraits — LudovicCreator · 2026-10-23
- AI-generated fantasy short 'The Sky Doesn't Kneel' stuns with dragon-rider visuals — scenemixer_official · 2026-10-23
- Stunning AI-generated Dark Knight images nail foggy Batman atmosphere — egeberkina · 2026-10-23
- AI video now pulls off orbiting, diving camera moves around a speeding bicycle — umesh_ai · 2026-10-23
- AI art showcase: gothic winged warrior "Azrael-42" with porcelain-crack armor — LudovicCreator · 2026-10-23