bug: candle-paddleocr-vl emits LaTeX markup and CJK glyphs for a handwritten signature scribble
Author: tobocop2Created Sep 18, 2026Updated Sep 19, 2026
LabelsbugP2
Description
candle-paddleocr-vl emits LaTeX markup and CJK glyphs for a handwritten signature scribble at the top of a page, in every run on that page. The rest of the page is read correctly and matches tesseract line for line within one percent of characters. Low severity: it is noise on a non-text region, but a consumer sees markup and foreign glyphs where there is nothing to read.
Expected: a scribble yields nothing, or a marker the caller can filter.
Steps to reproduce
- Run
candle-paddleocr-vlon a page image with a handwritten signature above printed text. - Read the first line of the output.
Relevant files and configuration
xberg v1.2.4, release build with candle-cuda, NVIDIA L4, 2026-09-18. About 23 s per page, 8.9 GB of GPU memory.
Source: xberg-io/xberg