Instructions to use dbcccc/TypLens with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use dbcccc/TypLens with Transformers:
# Use a pipeline as a high-level helper # Warning: Pipeline type "image-to-text" is no longer supported in transformers v5. # You must load the model directly (see below) or downgrade to v4.x with: # 'pip install "transformers<5.0.0' from transformers import pipeline pipe = pipeline("image-to-text", model="dbcccc/TypLens")# Load model directly from transformers import AutoTokenizer, AutoModelForMultimodalLM tokenizer = AutoTokenizer.from_pretrained("dbcccc/TypLens") model = AutoModelForMultimodalLM.from_pretrained("dbcccc/TypLens", device_map="auto") - Notebooks
- Google Colab
- Kaggle
License and attribution
Released weights and metadata
TypLens-V1 is a fine-tuned derivative of Pix2Text-MFR-1.5, not the original random-initialized IBEM-im2typst Phase10 model. The FP32 and INT8 exports are variants of the same fine-tuned checkpoint. Copyright (c) 2026 dbcccc.
The project contributions to these weights and metadata are licensed under the MIT License, to the extent the contributors hold applicable licensable rights. Upstream rights and notices are retained. Keep LICENSE, this notice, and the licenses directory with redistributed model files. The MIT grant does not require publication of training code. No training source is included.
Pretrained parameters
- Pix2Text-MFR-1.5 by BreezeDeus: model card, pinned source revision. The pinned model metadata declares MIT. The associated project's notice is Copyright (c) 2022 BreezeDeus; retained in licenses/PIX2TEXT-MIT.txt. Upstream license.
- The upstream model card attributes its initialization and architecture to Microsoft's TrOCR. The corresponding Microsoft MIT notice is retained in licenses/TROCR-MIT.txt. TrOCR project, license.
Changes include native Typst output targets and vocabulary, fine-tuning, screenshot augmentation, a cached ONNX decoder export, and an optional INT8 weight export.
Training-data attribution (datasets are not redistributed)
- IBEM, Dan Anitei, Joan Andreu Sanchez, and Jose Miguel Benedi: Zenodo record 7963703, CC BY 4.0. Formula crops and annotations were used to derive native Typst supervision.
- UniMER Dataset, Bin Wang, Zhuangcheng Gu, Chao Xu, Bo Zhang, Botian Shi, and Conghui He: dataset card, source revision 2343ddd963290469da36ca83e3a56c66e068add9. The dataset card declares Apache-2.0 and identifies component sources including Pix2tex, arXiv, CROHME and HME100K. It also gives a separate copyright-related download instruction for HME100K. The selected training images were not all traced to component-level rights. This release does not represent that every component image has a uniform license or provide a separate commercial-rights clearance for those source materials.
No dataset images, original annotations, or source PDFs are included. The model license does not relicense those materials. Runtime libraries, Typst, MiTeX, and ONNX Runtime binaries are also not included and retain their own licenses.