form-field-v1-small · commercial
A compact, high-accuracy form-field detector — locates Text, Choice (checkbox/radio), and Signature
widgets on empty, filled, and handwritten document pages. 8.9M params, ~36 MB; a strong accuracy/size
balance for on-device and edge GPU.
The runnable weights are commercial and are not downloadable from this repository — this public page is the
product specification and scorecard. Prefer a fully open model? See
form-field-v1-nano (Apache-2.0). Need the highest
accuracy? See form-field-v1-medium.
- 🎯 Try it: form-field-v1-demo
- 🏆 Leaderboard: form-field-v1-leaderboard
- 📊 Benchmark: form-field-v1-benchmark
Results
COCO mAP50-95 on the form-field-v1-benchmark,
one shared pycocotools scorer across all models.
| Model | Empty | Filled | Handwritten |
|---|---|---|---|
| FFDetr (baseline) | 0.461 | 0.170 | 0.225 |
| form-field-v1-small (this model) | 0.456 | 0.716 | 0.680 |
| FFDNet-L (baseline) | 0.373 | 0.285 | 0.285 |
Per-class on empty: Text 0.550 / Choice 0.556 / Signature 0.263. Throughput: ~11 pages/sec on a laptop CPU (ONNX Runtime). Full breakdowns on the leaderboard.
Output
[ { "box": [82, 164, 336, 41], "type": "Text", "score": 0.91 } ]
box is [x, y, w, h] in page pixels; type ∈ Text, ChoiceButton, Signature. Served as a single-file
ONNX graph (fp32 + fp16), verified to reproduce the benchmark (fp32 mAP 0.6161 / fp16 0.6134 over all variants).
Intended use and limits
- Empty, filled, and handwritten digital/scanned forms are all supported.
- On-device / edge GPU target; ~11 pages/sec CPU, faster on GPU.
License & data
Weights are offered under a commercial Nutrient license. Evaluation is public and reproducible via
form-field-v1-benchmark. The FFDetr
baseline is FF-DETR / CommonForms by jbarrow, Apache-2.0.
📩 Get access
form-field-v1-smallis commercial and its weights are not downloadable here. To run it on-prem — contact Nutrient: nutrient.io/contact-sales.
About the author
This project is maintained and funded by Nutrient - The deterministic document infrastructure enterprises run their highest-stakes workflows on: replayable output, clear exceptions, and full audit trails on the messy, regulated documents where AI alone breaks.