| license: apache-2.0 | |
| base_model: kha-white/manga-ocr-base | |
| language: | |
| - ja | |
| pipeline_tag: image-to-text | |
| tags: | |
| - onnx | |
| - vision-encoder-decoder | |
| - image-to-text | |
| - manga-ocr | |
| # manga-ocr (repackaged, full-precision) | |
| Full-precision manga-ocr-base repackaged into a single self-consistent repo for | |
| ONNX inference — matched weights **and** tokenizer in one place. | |
| - **Weights** (`encoder_model.onnx` 343 MB, `decoder_model.onnx` 117 MB): the | |
| full-precision, **non-merged** ONNX export (768-dim ViT encoder) from | |
| [`onnx-community/manga-ocr-base-ONNX`](https://huggingface.co/onnx-community/manga-ocr-base-ONNX), | |
| itself a transformers.js export of | |
| [`kha-white/manga-ocr-base`](https://huggingface.co/kha-white/manga-ocr-base). | |
| - **Tokenizer** (`tokenizer.json`, `vocab.txt`, 6144-token WordPiece): verified | |
| id→token-identical to the weights' lineage. | |
| I/O: encoder takes `pixel_values` → hidden states; decoder takes `input_ids` + | |
| `encoder_hidden_states` → logits (no KV-cache branch). | |