Introduction

You know the feeling. The lighting is perfect, the composition sings, and then the shop sign says "COFEFE" or one Chinese character melts into scribbles. PixText, a new open-weights model from the ATH-MaaS team, fixes exactly that. You draw a box around the broken text, type what it should say, and it repaints only that patch while every other pixel stays the same. It's an image-editing model under Apache-2.0, the weights went up on Hugging Face on October 22, and a ComfyUI custom node landed in the Comfy Registry on October 23 (11:13 MYT). You can run it today on a local NVIDIA GPU.

What Shipped

  • Two checkpoints: PixText-MeanFlow (4 sampling steps, the quick default) and PixText-Teacher (25 steps, a little higher quality). The MeanFlow model.safetensors is about 1.9 GB.
  • No extra baggage: MeanFlow needs no separate text encoder, OCR model, VAE or teacher. Each package ships its own fonts (Alibaba PuHuiTi and Go Noto CJK) plus SHA-256 checksums, so Latin and CJK text both render without system fonts.
  • Pixel-space editing: instead of round-tripping through a VAE, which can smudge thin strokes, PixText generates the target regions directly in pixel space. Crops snap to 16-pixel patches, and pixels outside the edited footprint are guaranteed to match your original image.
  • Batch fixes: up to 12 text targets per pass, generated jointly, on images of any size.
  • License: Apache-2.0 for the code and weights, plus an open benchmark set, TextRefinementBench.

The Numbers

The paper, "Efficient Scene-Text Refinement via Pixel-Space Grouped Image Generation" (arXiv 2610.18284), reports results on its new Text Refinement Bench:

  • Teacher: 83.06% glyph pass rate and 94.66% exact target-character OCR accuracy.
  • Strongest baseline it compared against: 36.77% and 52.31%.
  • MeanFlow (4 steps): keeps an 81.84% glyph pass rate at under one GPU-second of amortized generation cost per image.

Those are the authors' own numbers on their own benchmark, so treat them as a promising first read rather than an independent arena result.

Run It in ComfyUI

The custom node gives you a Load PixText Model node, a visual Region Editor for drawing boxes and typing the target text, and Compare Regions for before and after views. There's also an English workflow with Detect Words, which finds word boxes for you. Install it from your ComfyUI folder:

cd ComfyUI/custom_nodes
git clone https://github.com/ATH-MaaS/PixText.git
python -m pip install -r PixText/requirements.txt

Or use the Registry with comfy node install pixtext. Then download a full model package from Hugging Face into ComfyUI/models/pixtext/PixText-MeanFlow/, restart ComfyUI and open Templates → PixText. The Registry lists Linux with an NVIDIA CUDA GPU and ComfyUI 0.42.0 or newer. Version 0.1.3, posted October 24, moved the optional OCR download helper out of the Registry package. The nodes themselves didn't change.

For the English auto-detect workflow, add the optional word detector once:

cd ComfyUI/custom_nodes/PixText
python -m pip install -r requirements-ocr.txt
python scripts/download_ocr_models.py --output ../../models/pixtext_ocr

Run It in Python

You'll need Python 3.11 or newer, a CUDA build of PyTorch 2.8 that supports FlexAttention, and HarfBuzz (libharfbuzz0 on Debian or Ubuntu) so Pillow can shape text.

pip install "git+https://github.com/ATH-MaaS/PixText.git"

from pixtext import PixTextPipeline, TextEdit

pipeline = PixTextPipeline.from_pretrained("models/PixText-MeanFlow")
result = pipeline.edit(
    "poster.png",
    edits=[TextEdit(bbox=(120, 80, 360, 125), text="Night Market")],
    seed=42,
)
result.image.save("poster-fixed.png")

Boxes use original-image pixel coordinates (x0, y0, x1, y1). Give it one character or one line of text per box. A pixtext command-line tool does the same job from an edits.json file.

Tips for Clean Fixes

  • It's a repair tool, not a typesetter. PixText shines when the layout is already right and a few letters or characters went wrong. It won't design a new headline for you.
  • Box tightly. The edit can spill slightly past your box because of the 16-pixel patch grid, so leave a little breathing room around nearby details you care about.
  • Start with MeanFlow. If a tricky glyph still looks off, switch to Teacher, or rerun with a different seed.

Conclusion

Garbled text has been the last annoying mile of AI posters, menus and thumbnails. PixText turns it into a small, local, open fix that leaves the rest of your image alone. Grab the weights from the PixText model card, the code from the PixText GitHub repo, and the node from the Comfy Registry listing. Then go rescue that almost-perfect render sitting in your outputs folder.

—Aurelia ♡