Efficient Scene-Text Refinement via Pixel-Space Grouped Image Generation

Efficient Scene-Text Refinement via Pixel-Space Grouped Image Generation. By Yaoli Liu and six others, from Zhejiang University, Alibaba Group, and one other institution. Published on the thirteenth of October. No code repository is linked to it.

Generated images often get the layout and the visual content right, yet small text and complex characters still come out with local errors, and that limits how readable and usable the images are. Fixing scene text asks for three things at once: accuracy, preservation of everything else, and efficiency. Existing methods built on variational autoencoders can lose fine strokes when they reconstruct the image. Editing the full image instead can shift the background colors.

PixText reframes local text refinement as grouped image generation, carried out in pixel space rather than through an autoencoder. Each of the two failure modes above gets a direct answer. Staying in pixels avoids the reconstruction step where strokes were lost, and generating locally avoids touching the background at all.

Conditioned on the original image, PixText generates only the target regions and leaves unedited content as it was. It uses MeanFlow so that generation takes four steps. To measure the task, the authors also introduce Text Refinement Bench, which evaluates both text correction and image preservation in generated images. Scoring both matters, since a method could fix the text perfectly and still damage everything around it.

On that benchmark, PixText achieves an eighty-three point oh six percent glyph pass rate. Its exact target-character OCR accuracy is ninety-four point six six percent. The strongest baseline reaches thirty-six point seven seven percent on the first measure, and fifty-two point three one percent on the second.

The four-step MeanFlow variant keeps an eighty-one point eight four percent pass rate, at an amortized cost of less than one GPU-second per image.

Two choices carry the design. Working in pixels keeps the strokes, and regenerating only the broken regions keeps the rest of the picture where it was.