Baidu's PP-DocLayoutV3 Reads Curved and Skewed Docs With 33M Parameters
PP-DocLayoutV3 tackles skewed, curved, and warped documents with polygon boxes and reading order in a single 33M-parameter model.
Takeaways
- PaddlePaddle released PP-DocLayoutV3, a 33M-parameter layout detector for warped and skewed documents
- Predicts polygon boxes and reading order end-to-end in a single forward pass
- Built on RT-DETR with a mask-based detection head replacing rectangular boxes
- Powers layout analysis inside PaddleOCR-VL-1.5 and GLM-OCR pipelines
- Apache 2.0 licensed, safetensors format, works with Hugging Face Transformers directly
- Handles skew, curving, screen photos, and lighting variation better than PP-DocLayoutV2
Layout analysis is the unglamorous step that turns a photo of a document into something a language model can actually parse. Baidu's PaddleOCR team just released PP-DocLayoutV3, a compact detector that predicts polygon-shaped boxes and reading order in a single forward pass, built for the real-world messiness of phone-captured pages, curved book spreads, and screen photos.
The model powers layout analysis inside PaddleOCR-VL-1.5 and GLM-OCR. It ships under Apache 2.0 with safetensors weights on Hugging Face, weighs in at 33.3M parameters, and has already crossed 895K downloads on the safetensors repo alone.
What V3 changes
Prior versions of PP-DocLayout used a two-step recipe: an RT-DETR detector predicts rectangular boxes, then a pointer network with six transformer layers sorts them into reading order. That works for clean scans but breaks when a page is skewed, folded, or shot at an angle.
V3 replaces the detection head with an instance segmentation framework and folds reading order prediction into the same end-to-end transformer. Instead of drawing an axis-aligned rectangle around a paragraph rotated 15 degrees, the model outputs a tight polygon hugging the actual text region, and handles classification, localization, masking, and ordering in a single forward pass. Cascading errors from a separate ordering stage largely disappear as a result.