Zhai et al. at Google introduced SigLIP, replacing CLIP's global batch-wide softmax contrastive loss with a pairwise sigmoid loss. Removing the all-gather normalization bottleneck improved small- and large-batch scaling, making SigLIP a preferred vision encoder for modern VLMs.
CLIP
Also known as: Contrastive Language-Image Pre-training · OpenCLIP · SigLIP
Plain English
A dual-encoder model that learns to map images and text descriptions into the same shared mathematical space, forming the bridge that lets image generators understand prompts and vision-language models see.
Technical Overview
Introduced by OpenAI, CLIP jointly trains an image encoder (ViT or ResNet) and a text Transformer with a symmetric contrastive loss over 400M image-text pairs so matching pairs align in cosine similarity. Frozen CLIP (and successors like OpenCLIP and Google's SigLIP) became the standard text-conditioning encoder for Stable Diffusion and FLUX as well as the vision backbone for modular VLMs.
Chronological Entries (2)
Radford et al. at OpenAI published CLIP, showing that contrastive pretraining on 400 million web image-text pairs produces zero-shot visual representations rivaling supervised ImageNet models—and providing the text-image embedding space that later powered text-to-image diffusion.
Direct Relations (5)
- Created byOpenAI
- UsesTransformer
- EnablesMultimodal & Vision-Language Models
- EnablesStable Diffusion
- EnablesFLUX
Referenced By (0)
No incoming references yet.