UVDoc ONNX grid model

A grid-output derivative of PaddlePaddle/UVDoc_onnx, prepared for browser-local document flattening in Baimiao PDF. This is not a newly trained model and is not the full upstream image-output graph.

Artifact

  • File: uvdoc-grid-v1.onnx
  • Size: 31,682,817 bytes
  • SHA-256: a3b1cbf9ca5216c39cd53cb63af8fbe91414e74d0f0adb1a6f1ba38f56979ae6
  • Upstream input model SHA-256: 54cab30dc2cf347b4f3d6971c833b1c3d84f5ef17280f1f6d15796bff20e63d6

Changes and usage

export-grid.py exposes the coarse two-channel sampling grid before the final Resize operation and removes the downstream image-resampling graph. No weights were retrained. Input is a float32 NCHW tensor shaped [1, 3, 712, 488], BGR channels scaled to [0, 1]. Output contains normalized x/y source coordinates, using align-corners semantics. The caller must interpolate the grid and resample the original image; model output is not a corrected image or OCR text.

The Baimiao browser adapter follows the grid export used by crop-dewarp-scan-erase at commit e68c4d8ab0ce5f00245d5316d664ada858afb8a5. Additional text, table and page-edge correction code is outside this model repository.

Attribution and license

Upstream model and model card: PaddlePaddle/PaddleOCR contributors, PaddlePaddle/UVDoc_onnx. Upstream declares Apache-2.0; see LICENSE and the preserved UPSTREAM-MODEL-CARD.md. This derivative is distributed under Apache-2.0. Modification: grid-output export for browser inference, 2026-11-17.

Both kittypdf mirrors distribute the same bytes. Applications should verify the pinned size and SHA-256 before loading.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for kittypdf/UVDoc-ONNX

Quantized
(1)
this model