README.md

---
license: apache-2.0
pipeline_tag: image-to-text
---
<div align="center">

<h1>Tokenize Anything via Prompting</h1>

[Ting Pan](https://github.com/PhyscalX/)<sup>1,2*</sup>, &nbsp; [Lulu Tang](https://github.com/lulutang0608)<sup>2*</sup>, &nbsp; [Xinlong Wang](https://www.xloong.wang/)<sup>2¶</sup>, &nbsp; [Shiguang Shan](https://scholar.google.com/citations?user=Vkzd7MIAAAAJ&hl=en)<sup>1</sup>

<sup>1</sup>[ICT-CAS](http://english.ict.cas.cn/), &nbsp; <sup>2</sup>[BAAI](https://www.baai.ac.cn/english.html)<br>
<sup>*</sup> Equal Contribution, <sup>¶</sup>Project Lead

[[`Paper`](https://arxiv.org/pdf/2312.09128.pdf)] [[`🤗 Demo`](https://huggingface.co./spaces/BAAI/tokenize-anything)]

</div>

We present **T**okenize **A**nything via **P**rompting, a unified and promptable model capable of simultaneously segmenting, recognizing, and captioning arbitrary regions, with flexible visual prompts (point, box and sketch). The model is trained with exhaustive segmentation masks sourced from SA-1B, coupled with semantic priors from a pre-trained EVA-CLIP with 5 billion parameters.

## Installation

See [Github Page](https://github.com/baaivision/tokenize-anything).

## Models

### Model weights

#### V1.1 Release Notes

- Three versions of the model are available with different image encoders.
- Use a longer pre-training and fine-tuning schedule (improved segmentation and caption performance).
- Apply weight decay for all bias parameters (avoid FP16 overflow in QK matmul).
- Sample point prompts from predicted mask instead of GT box during VG training.

| Model | Description | Schedule | MD5 | Weights |
| ----- | ------------| ------ | ----| ------ |
| **tap_vit_h** | ViT-H TAP v1.1 model | (100% SA-1B, 180k), (VG, 50ep) | 4bdfb9 | [🤗 HF link](https://huggingface.co./BAAI/tokenize-anything/blob/main/models/tap_vit_h_v1_1.pkl) |
| **tap_vit_l** | ViT-L TAP v1.1 model | (100% SA-1B, 180k), (VG, 50ep) | c1d41f | [🤗 HF link](https://huggingface.co./BAAI/tokenize-anything/blob/main/models/tap_vit_l_v1_1.pkl) |
| **tap_vit_b** | ViT-B TAP v1.1 model | (100% SA-1B, 180k), (VG, 50ep) | 707f80 | [🤗 HF link](https://huggingface.co./BAAI/tokenize-anything/blob/main/models/tap_vit_b_v1_1.pkl) |

#### V1.0 Release Notes

- Two versions of the model are available with different image encoders.
- Original paper results.

| Model | Description | Schedule | MD5 | Weights |
| ----- | ------------| ------ | ----| ------ |
| **tap_vit_l** | ViT-L TAP v1.0 model | (50% SA-1B, 90k), (VG, 25ep) | 03f8ec | [🤗 HF link](https://huggingface.co./BAAI/tokenize-anything/blob/main/models/tap_vit_l_v1_0.pkl) |
| **tap_vit_b** | ViT-B TAP v1.0 model | (50% SA-1B, 90k), (VG, 25ep) | b45cbf | [🤗 HF link](https://huggingface.co./BAAI/tokenize-anything/blob/main/models/tap_vit_b_v1_0.pkl) |

### Concept weights

***Note***: You can generate these weights following the [Concept Guide](https://github.com/baaivision/tokenize-anything/blob/main/notebooks/concept.ipynb).

| Concept | Description | Weights |
| ------- | ------------| ------ |
| **Merged-2560** | Merged concepts | [🤗 HF link](https://huggingface.co./BAAI/tokenize-anything/blob/main/concepts/merged_2560.pkl) |
| **LVIS-1203**   | LVIS concepts | [🤗 HF link](https://huggingface.co./BAAI/tokenize-anything/blob/main/concepts/lvis_1203.pkl) |
| **COCO-80**   | COCO concepts  | [🤗 HF link](https://huggingface.co./BAAI/tokenize-anything/blob/main/concepts/coco_80.pkl) |

## License
[Apache License 2.0](LICENSE)

## Citation

```
@article{pan2023tap,
  title={Tokenize Anything via Prompting},
  author={Pan, Ting and Tang, Lulu and Wang, Xinlong and Shan, Shiguang},
  journal={arXiv preprint arXiv:2312.09128},
  year={2023}
}
```

## Acknowledgement

We thank the repositories: [SAM](https://github.com/facebookresearch/segment-anything), [EVA](https://github.com/baaivision/EVA), [LLaMA](https://github.com/facebookresearch/llama), [FlashAttention](https://github.com/Dao-AILab/flash-attention), [Gradio](https://github.com/gradio-app/gradio), [Detectron2](https://github.com/facebookresearch/detectron2) and [CodeWithGPU](https://github.com/seetacloud/codewithgpu).