Ships the de-htr-web-v2 model as the default recognizer, with v1 selectable in-app for comparison.
Highlights
- v2 model default: 8.01% CER / 32.3% WER on the sealed writer-disjoint ScaDS test split (v1: 8.61% / 33.9%), greedy CTC at batch 1 — the deployment configuration.
- Config-driven model contracts (
format_version): v1 keeps its RGB + external ImageNet normalization path byte-identical; v2 uses 1-channel grayscale, [0,1] scaling, in-model normalization, raw-logits output, 112-symbol alphabet, max width 1024. - Model picker UI — switch v1/v2 live and re-run on the loaded image.
- Fix: v2 emits batch-major
[1, T, C](v1 is time-major[T, 1, C]); decode verified bit-exact against the Python reference pipeline.
Artifacts
- v2 int8 ONNX: 10,396,365 bytes, SHA-256
b576a0a1281b9be46b2574028b75575e041b7d7cb650f063886e733467cc1499 - Model + card + config: https://huggingface.co/naeyn/de-htr-web-v2
- Production demo: https://de-htr-web.vercel.app
- Training/ablation research record: tiny-htr project (ablation summary published with the model card)
Scope
Modern German handwriting, one pre-cropped line; no page segmentation, no historical scripts (Kurrent/Sütterlin). Weights Apache-2.0; training-data attribution in the model repo's NOTICE-training-data.md.