CLIP ViT-B/16 LAION-2B

laion/CLIP-ViT-B-16-laion2B-s34B-b88K is a zero-shot image classification model that performs contrastive language-image pre-training (CLIP) using a ViT-B/16 vision encoder and a text encoder, trained on the 2-billion-sample English subset of LAION-5B with OpenCLIP. The model is hosted as a managed API on gigarouter, providing OpenAI-compatible endpoints for zero-shot inference.

Key Capabilities

• Zero-shot image classification: assign labels to images without task-specific fine-tuning by matching image embeddings to text prompt embeddings.
• Image and text retrieval: rank images by text queries or vice versa using cosine similarity between embeddings.
• Downstream adaptability: can be fine-tuned for image classification, used as a frozen feature extractor for linear probing, or employed for image generation guidance and conditioning.

Training and Data

The model was trained by Mehdi Cherti on the JUWELS Booster supercomputer at Jülich Supercomputing Centre. Training data is the English subset of LAION-5B, an uncurated dataset of 2 billion image-text pairs crawled from public internet. The dataset is intended for research; users should be aware of potential harmful content.

Benchmark Results

The model achieves a 70.2% zero-shot top-1 accuracy on ImageNet-1k. Evaluation was performed using the LAION CLIP Benchmark suite on VTAB+ (Visual Task Adaptation Benchmark with additional robustness datasets) for classification and on COCO and Flickr for retrieval. Extended benchmark results are available in the CLIP benchmark repository.

Usage Notes

As per the original OpenAI CLIP model card, this model is a research output intended for research communities. Deployed or unconstrained production use is not recommended without thorough domain-specific testing. Use in surveillance or facial recognition is out of scope. The model is designed for English-language inputs only.