| # TypLens V1.1 |
|
|
| V1.1 introduces a newly trained single-channel grayscale model and matched |
| training/browser preprocessing for printed formula images. It uses the final |
| 33,544-step checkpoint and generates Typst directly. |
|
|
| - Full FP32 ONNX: **117.2 MB** of model weights. |
| - Compact dynamic INT8 ONNX: **33.1 MB**, 71.7% smaller. |
| - Native FP32 safetensors, tokenizer and configuration are available on |
| [Hugging Face](https://huggingface.co/dbcccc/TypLens). |
| - Both ONNX variants have completed desktop Edge/WASM smoke checks. |
|
|
| **Migration:** the input changes from three RGB channels to one luminance channel. |
| Adopt the preprocessing in INFERENCE.md and replace the encoder and decoder |
| together. Old V1 RGB preprocessing cannot be used with these weights. |
|
|
| The server FP32 checkpoint received 127/135 content passes on a reused development |
| set under assistant review. This is not independent accuracy or a compact-model |
| benchmark. The model card explains the scope and remaining errors. |
|
|
| Attach `typlens-v1.1-fp32-onnx.zip` and `typlens-v1.1-int8-onnx.zip` to this release. |
| Each asset includes its model pair, token bytes, preprocessing contract and license |
| notices. The release contains no training code. The MIT license and upstream |
| attribution are retained. |
|
|
