open_clip

Why we included this project

OpenCLIP is the open-source CLIP implementation most teams build on when they need contrastive vision-language training in PyTorch. It learns a shared embedding space for images and text, which covers zero-shot classification, cross-modal retrieval, and semantic search, and the repo includes both the code to train models and a large set of pretrained checkpoints that load in a few lines. Training scales from fine-tuning on a single GPU to full clusters, with distributed and FSDP2 support. The main branch has grown well beyond the original model, adding variable-resolution image towers, audio-text variants, and newer decoder-style families, while the v3 branch keeps a stable API for teams that prefer not to chase experimental changes. If you would otherwise reimplement contrastive pretraining from scratch, this is a solid, reproducible starting point.