license: apache-2.0
library_name: flattener
pipeline_tag: image-to-image
language:
- en
- ko
tags:
- document-scanning
- document-dewarping
- onnx
- pytorch
Flattener
Version 1.1.0 of the models for Flattener. Flattener turns document photos and open books into flat, upright scans. Processing runs locally.
Files
| File | Purpose |
|---|---|
dewarp.onnx, dewarp.pt |
Correct page perspective and curvature |
gate.onnx, gate.pt |
Detect the page region |
orient.onnx, orient.pt |
Predict text orientation |
manifest.json |
Load the three ONNX models together |
bundle.json |
Load the three PyTorch checkpoints together |
The ONNX models are exported from the PyTorch checkpoints and hold the same weights.
The scanner runs them with ONNX Runtime on the CPU by default.
The PyTorch checkpoints run on a CUDA GPU and are the starting point for training.
manifest.json and the orientation checkpoint include the calibrated orientation temperature and threshold.
Use
uvx flattener-scan -i photo.jpg scan.jpg
The first scan downloads the ONNX models at revision 1.1.0.
Later scans reuse the local Hugging Face cache, including offline.
With PyTorch installed and a CUDA GPU present (uvx --with torch flattener-scan ...), the scanner downloads and uses the PyTorch checkpoints instead.
Download the files to a chosen directory with:
hf download DevWorld/flattener manifest.json dewarp.onnx gate.onnx orient.onnx --revision 1.1.0 --local-dir models
Then pass --bundle models/manifest.json to the scanner.
Training and use
The models were trained from scratch using SyntheticDoc renders, UVDoc paper captures, SmartDoc frames and generated pages. The dewarp model blends a clean-page checkpoint with a checkpoint trained on additional creases and folds. The page detection and orientation models include generated Korean and English documents. See the source repository for training, evaluation and browser export tools.
Pages cut off by the photo, heavy occlusion and low-resolution text can produce incomplete scans or uncertain orientation. The scanner reports these conditions and supports manual page and orientation adjustments.
License
These model weights are licensed under Apache-2.0.