| license: other | |
| tags: | |
| - onnx | |
| - manga | |
| - text-detection | |
| - inpainting | |
| - webgpu | |
| # Tara Models | |
| ONNX models for [Tara](https://github.com/QuatZo/tara) - browser-native visual document translation. | |
| ## Models | |
| | Model | File | Size | Source | Changes | | |
| |-------|------|------|--------|---------| | |
| | RT-DETR-v2 (detector) | detector.onnx | 161 MB | [ogkalu/comic-text-and-bubble-detector](https://huggingface.co/ogkalu/comic-text-and-bubble-detector) | Patched AveragePool ceil_mode=0 for WebGPU compatibility | | |
| | LaMa (inpainter) | lama-manga-dynamic.onnx | 197 MB | [ogkalu/lama-manga-onnx-dynamic](https://huggingface.co/ogkalu/lama-manga-onnx-dynamic) | No changes, original model | | |
| ## Why this repo? | |
| The original detector.onnx uses AveragePool with ceil_mode=1, which onnxruntime-web's WebGPU backend does not support. This repo hosts a patched version where ceil_mode is set to 0. Since all AveragePool nodes use kernel=2x2 stride=2 on even-dimension inputs (640x640), this change has no effect on output. | |
| The LaMa inpainter works as-is and is included here for convenience (single download source). | |
| ## Usage | |
| These models are loaded by Tara via the browser Cache API. On first visit, the app downloads both models (~360 MB total) and caches them locally. Subsequent visits load from cache instantly. | |
| Direct download URLs: | |
| - https://huggingface.co/QuatZo/tara-models/resolve/main/detector.onnx | |
| - https://huggingface.co/QuatZo/tara-models/resolve/main/lama-manga-dynamic.onnx | |
| ## License | |
| Original models by [ogkalu](https://huggingface.co/ogkalu). See the source repositories for license information. | |