Is PixText state of the art?Pending benchmarks.Announced Oct 22, 2026; no independent results yet.
A model that fixes small text inside generated images directly in pixel space: give it an image, the regions to edit and the target text, and it redraws only those regions, leaving other pixels unchanged. Two Apache 2.0 checkpoints on ModelScope: PixText-MeanFlow runs in 4 steps, PixText-Teacher in 25.
The details so far
- Lab
- ATH-MaaS
- Announced
- Oct 22, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- ~0.95B per checkpoint
- Lab
- ATH-MaaS on ModelScope; paper authors from Zhejiang University and Alibaba Group
- Weights
- Apache 2.0 on ModelScope (ATH-MaaS/PixText): PixText-MeanFlow and PixText-Teacher, BF16
- Sampling
- MeanFlow: 4-step interval Euler; Teacher: 25-step Euler
- Design
- Pixel-space generation with no VAE or text encoder; target text is given as rendered glyph images
- API
- None announced
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on its own Text Refinement Bench, 83.06% glyph pass and 94.66% exact target-character OCR accuracy, versus 36.77% and 52.31% for FLUX-Text; the 4-step MeanFlow variant keeps 81.84% at under one GPU-second per image.
Sources
REPOPixText on ModelScopemodelscope.ai/models/ATH-MaaS/PixText · Oct 22, 2026PAPEREfficient Scene-Text Refinement via Pixel-Space Grouped Image Generation (arXiv 2610.18284)arxiv.org/abs/2610.18284 · Oct 22, 2026LABModelScope announcement on Xx.com/ModelScope2022/status/2113204928076247522 · Oct 22, 2026Tracking continues until an independent board lists it.