Sigmoid Loss for Language Image Pre-Training (SigLIP)

Softmax/InfoNCE contrastive loss requires a global view of all in-batch similarities, which caps the usable batch size and makes large-scale pre-training expensive.

Replaces the softmax-normalised contrastive loss with a PAIRWISE sigmoid loss computed only on image-text pairs, so no batch-wide normalisation is needed. This decouples loss quality from batch size. Combined with Locked-image Tuning (freeze the image tower, precompute embeddings) this made very large batches affordable.

84.5% ImageNet zero-shot accuracy in two days using only FOUR TPUv4 chips with Locked-image Tuning. That is a training-cost claim, not an inference speedup - it is recorded here because it is the clearest hardware+baseline statement in the CLIP-lineage literature. The paper also pushes batch size to one million and finds diminishing returns past ~32k.

ICCV 2023 Oral (arXiv comment; crossref confirms the ICCV 2023 container). SigLIP is the encoder behind PaliGemma and Gemma-style vision towers and is a standard CLIP replacement wherever training batch size was the binding constraint. Serving behaviour is identical to CLIP - frozen tower, one pass per image - so the inference cost profile follows vision encoder size and input resolution, not the loss. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: iccv has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.