SigLIP-Base
A vision-language foundation model based on image-text contrastive learning that maps images and text into a shared embedding space and supports zero-shot image classification. It has a relatively large parameter count and computational cost.
Timeline (last 90 days)
Pasma's comparison is a deployment comparison, not an equal-training-compute ablation, because Pasma is the result of task fine-tuning and distillation, whereas SigLIP-Base remains in a zero-shot state.
On a held-out combined set of 1,103 real photos, Pasma-Small achieved 89.12% Top-1 accuracy for clothing type, versus 80.24% for SigLIP-Base zero-shot.
On a subset containing only 249 home/catalog images, Pasma's Top-1 accuracy was 83.53% versus 87.15% for SigLIP-Base, reversing the ranking.
SigLIP-Base is a vision-language contrastive learning model that, through training on large-scale image-text pairs, maps images and text into the same embedding space, enabling classification in zero-shot inference by matching text descriptions with images.
All Facts (4)
Pasma's comparison is a deployment comparison, not an equal-training-compute ablation, because Pasma is the result of task fine-tuning and distillation, whereas SigLIP-Base remains in a zero-shot state.
50%UnverifiedOn a held-out combined set of 1,103 real photos, Pasma-Small achieved 89.12% Top-1 accuracy for clothing type, versus 80.24% for SigLIP-Base zero-shot.
50%UnverifiedOn a subset containing only 249 home/catalog images, Pasma's Top-1 accuracy was 83.53% versus 87.15% for SigLIP-Base, reversing the ranking.
50%UnverifiedSigLIP-Base is a vision-language contrastive learning model that, through training on large-scale image-text pairs, maps images and text into the same embedding space, enabling classification in zero-shot inference by matching text descriptions with images.
50%