|
|
--- |
|
|
license: apache-2.0 |
|
|
library_name: open_clip |
|
|
pipeline_tag: image-feature-extraction |
|
|
--- |
|
|
|
|
|
# OpenVision 3: A Family of Unified Visual Encoder for Both Understanding and Generation |
|
|
|
|
|
OpenVision 3 is a family of advanced vision encoders that learn a single, unified visual representation serving both image understanding and image generation. |
|
|
|
|
|
Our core architecture is simple: we feed VAE-compressed image latents to a ViT encoder and train its output to support two complementary roles. First, the encoder output is passed to the ViT-VAE decoder to reconstruct the original image, encouraging the representation to capture generative structure. Second, the same representation is optimized with contrastive learning and image-captioning objectives, strengthening semantic features. |
|
|
|
|
|
- **Paper:** [OpenVision 3: A Family of Unified Visual Encoder for Both Understanding and Generation](https://huggingface.co/papers/2601.15369) |
|
|
- **Project Page:** [https://ucsc-vlaa.github.io/OpenVision3/](https://ucsc-vlaa.github.io/OpenVision3/) |
|
|
- **Repository:** [https://github.com/UCSC-VLAA/OpenVision](https://github.com/UCSC-VLAA/OpenVision) |
|
|
|
|
|
## Usage |
|
|
|
|
|
The OpenVision 3 vision encoders are compatible with the `open_clip` library. |
|
|
|
|
|
```python |
|
|
import torch |
|
|
import torch.nn.functional as F |
|
|
from urllib.request import urlopen |
|
|
from PIL import Image |
|
|
from open_clip import create_model_from_pretrained, get_tokenizer |
|
|
|
|
|
# Replace 'UCSC-VLAA/openvision3-model-id' with the actual repository ID |
|
|
model_id = 'hf-hub:UCSC-VLAA/openvision3-model-id' |
|
|
|
|
|
model, preprocess = create_model_from_pretrained(model_id) |
|
|
tokenizer = get_tokenizer(model_id) |
|
|
|
|
|
# Load and preprocess image |
|
|
image = Image.open(urlopen( |
|
|
'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png' |
|
|
)) |
|
|
image = preprocess(image).unsqueeze(0) |
|
|
|
|
|
# Prepare text queries |
|
|
text = tokenizer(["a diagram", "a dog", "a cat", "a beignet"], context_length=model.context_length) |
|
|
|
|
|
with torch.no_grad(), torch.cuda.amp.autocast(): |
|
|
image_features = model.encode_image(image) |
|
|
text_features = model.encode_text(text) |
|
|
|
|
|
image_features = F.normalize(image_features, dim=-1) |
|
|
text_features = F.normalize(text_features, dim=-1) |
|
|
|
|
|
text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1) |
|
|
|
|
|
print("Label probs:", text_probs) |
|
|
``` |
|
|
|
|
|
## Citation |
|
|
|
|
|
```bibtex |
|
|
@article{zhang2026openvision3, |
|
|
title = {OpenVision 3: A Family of Unified Visual Encoder for Both Understanding and Generation}, |
|
|
author = {Zhang, Letian and Ren, Sucheng and Liu, Yanqing and Li, Xianhang and Wang, Zeyu and Zhou, Yuyin and Yao, Huaxiu Zheng, Zeyu and Nie, Weili and Liu, Guilin and Yu, Zhiding and Xie, Cihang}, |
|
|
journal = {arXiv preprint arXiv:2601.15369}, |
|
|
year = {2026} |
|
|
} |
|
|
``` |
