Extract Text (PaddleOCR VL)

@paddlepaddle1 VersionExtract

PaddleOCR-VL document parsing. Each Image is parsed on the selected device with layout detection and vision-language recognition, emitting a joined String transcript or per-block regions with boxes. SOTA parser (PaddleOCR-VL-1.6, 0.9B); CUDA GPU required.

How it fits

Typical backends

Document to text with LLM post-processing over HTTP.

Live capture cropped then OCR'd.

Caveats

Hard constraintThis is a GPU vision-language model. The selected device must reach a CUDA GPU (`gpu:0`); CPU inference is impractical and unsupported by the upstream model.
I/O contractThe output is a oneof: a joined String transcript, or a list of regions each carrying `bounds` (a Rectangle or Polygon) and `text`, optionally with `confidence`. The downstream connection resolves which arm; regions follow reading order.
Parameter interactionThe pipeline version selects the model release (`v1.6` SOTA, `v1.5` prior). The maximum image side bounds memory and latency by downscaling oversized pages before parsing.
I/O contractAn empty Image returns the empty arm without invoking the model; weights are downloaded on first run and cached under the PaddleX model store.

Versions

2300290clatestdefault
Automated release
6/28/2026