- Who it helps
- Designers and developers correcting small labels or captions in generated images.
- Availability
- Code and model packages are public under Apache 2.0, with separate font terms. Inference requires an NVIDIA CUDA GPU and Python 3.11 or later; this is not a native Mac inference setup.
Give it a bounded correction
Supply the image, a pixel-coordinate box and the exact replacement text. The default MeanFlow package uses four updates. The model’s aligned crop can extend beyond the requested box, so inspect the surrounding image too.
Keep the test small
The accompanying guide creates a fictional COFFEE label and a separate KEEP label. Its helper checks whether changes leave an approved region. The helper cannot judge letter accuracy; you must read the final word yourself.
PixText implementation ↗PixText scene-text refinement paper ↗
What you can do
- Use the guide’s fictional input and retain the original, output and model metadata before adapting the process to an important image. Open reference ↗
Where the evidence stops
The original fixture and boundary helper were tested locally. PixText inference, reported speed and text accuracy were not tested; the creator post is not independent reproduction.
Sources read. Public evidence checked October 24, 2026 UTC. No third-party software, model inference or complete product workflow was executed.
Practical guide
PixText: repair one word and check what else changed →
A 512 × 192 input image, a corrected COFFEE label, the model metadata and a separate record of any changes outside the approved rectangle.
Sources & further reading
GitHub · implementation
PixText implementation ↗
The README documents image, box and target-text inputs, CUDA requirements and preservation outside actual edited footprints.
ATH-MaaS · page checked 2026-10-24
GitHub · implementation
PixText package dependencies ↗
The package declares version 0.1.2, Python 3.11+, Linux with NVIDIA CUDA, torch 2.8.0, torchvision 0.23.0 and Pillow 13.0.0. These pins define the guide environment; they have not been installed here.
ATH-MaaS · page checked 2026-10-24
Paper · paper
PixText scene-text refinement paper ↗
The October 13 paper defines the evaluation and reports author-measured accuracy and amortized GPU time.
Liu and colleagues · page checked 2026-10-24
Reddit · demonstration
PixText creator demonstration and discussion ↗
The maker links the same repository and model card. Visible post score was 277 with 24 comments; those counts do not establish independent replication.
Reddit · page checked 2026-10-24
277 net score · 24 comments · observed 2026-10-24 08:23 UTC
Web · implementation
PyTorch previous-version installation commands ↗
The official archive lists the torch 2.8.0 and torchvision 0.23.0 CUDA 12.8 wheels used in the documented setup. Installation was not executed.
PyTorch · page checked 2026-10-24
More sources reviewed (2)
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
PixText edits selected words in an existing image while preserving pixels outside its actual edit regions.