← sources

PixText: a small, fast model for fixing text in generated images

source: reddit (reddit:StableDiffusion)published 2026-10-23 09:09ingested 2026-10-24 02:38

retained text

I built PixText for a very specific frustration: an image looks great, most of the text is right, but a few mangled words or Chinese characters keep me from being happy with it.

VAE compression can lose fine details in small text regions, so I trained the model directly in pixel space. I also wanted the refiner to be small and fast: it's under 1B parameters (about 0.95B), needs no text encoder or VAE, and uses MeanFlow to refine text in just 4 sampling steps.

If you'd like to try it:

* [ComfyUI node on the Registry](https://registry.comfy.org/publishers/yaoliliu/nodes/pixtext)
* [Ready-to-use ComfyUI workflows](https://github.com/ATH-MaaS/PixText/tree/main/example_workflows)
* [Code and installation](https://github.com/ATH-MaaS/PixText#comfyui)
* [Model weights](https://huggingface.co/ATH-MaaS/PixText)

I'd love to hear what you think. Feedback and examples are very welcome! 🤗

document fetched 2026-10-24 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xd9186/pixtext_a_small_fast_model_for_fixing_text_in