| name | open-clip |
| description | OpenCLIP — open-source implementation of CLIP trained on LAION-5B/OpenCLIP datasets. Multi-head attention pooling, SigLIP loss variants, and wide model zoo (ViT, ConvNeXt, EVA). Community-driven. |
| tags | ["open-clip","multimodal","image-text","laion","zero-shot","embeddings","zorai"] |
Overview
OpenCLIP is an open-source reimplementation of CLIP trained on LAION-5B, LAION-400M, and DataComp. Provides larger and better architectures than the original: ViT-H/14, ConvNeXt, EVA-02, SigLIP. Full model transparency with flexible training customizations.
Installation
uv pip install open-clip-torch
Encoding Images and Text
import open_clip
import torch
from PIL import Image
model, _, preprocess = open_clip.create_model_and_transforms(
"ViT-H-14", pretrained="laion2b_s32b_b79k")
tokenizer = open_clip.get_tokenizer("ViT-H-14")
image = preprocess(Image.open("photo.jpg")).unsqueeze(0)
text = tokenizer(["a dog", "a cat", "a car"])
with torch.no_grad():
image_features = model.encode_image(image)
text_features = model.encode_text(text)
logits = (image_features @ text_features.T).softmax(dim=-1)
print(f"Predicted: class {logits.argmax().item()} with {logits.max():.2%}")
References