GitHub Repo
Transformers Fixes PaddleOCR-VL Position Interpolation Setting, Restoring Behavior Consistent With the Original Model
On November 5, Transformers merged a PaddleOCR-VL fix restoring the visual positional embedding interpolation alignment setting. Community tests show only a small difference in overall document recognition scores; the change mainly improves implementation consistency and reproducibility.
Hugging Face Transformers merged fix #50384 on November 5, changing interpolation_align_corners in PaddleOCR-VL’s visual embeddings from True back to False. This restores the bilinear interpolation behavior used by the original model. The change has landed on the main branch, but currently available information does not confirm which official package release will include it.
The issue occurs when visual positional embeddings must be resampled to match the input image grid. According to the reporter’s version comparison, version 5.12.1 used align_corners=False; starting with 5.13.0, a refactor of a shared visual helper function adopted coordinate rules equivalent to True. When the target patch grid differs from the grid learned by the model, the interpolated positional vectors change. A minimal reproduction that scales a 2×2 grid to 4×4 produced a maximum absolute difference of about 0.25 compared with the original setting. This difference comes from synthetic test vectors and should not be interpreted directly as a recognition error rate. Issue and reproduction code
This detail matters for document recognition deployments. PaddleOCR-VL’s core model is a 0.9B vision-language model that combines a NaViT-style dynamic-resolution visual encoder with ERNIE-4.5-0.3B to handle text, tables, formulas, and charts. For pipelines that need to support different page sizes, whether positional interpolation matches the original implementation is a variable to control when reproducing results across versions. Model documentation
However, the public evidence does not indicate a severe accuracy regression. The reporter used the same layout model through PaddleX to run OmniDocBench v1.5, reporting overall scores of 95.25 for False and 95.20 for True; the official vLLM reference result was 95.24. Both settings could also produce errors or duplicate output on difficult cases. These tests were provided by a contributor and are not sufficient to infer the impact across all documents. Engineering teams can first pin model weights and package versions, then compare outputs on their own Chinese tables, scanned pages, and documents of different sizes, while tracking inclusion in an official release. Evaluation discussion