SigLIP 2: sigmoid image–text alignment and masked-view self-distillation¶

简体中文

Fidelity: core mechanism. Runs sigmoid contrastive training and masked-view self-distillation, not multilingual web data, caption decoding, or dense localization.

Paper information¶

Field	Details
Paper	arXiv 2502.14786
Organization	Google DeepMind
First public release	2025-02-20 (arXiv v1)
Authors' code	Public implementation
Adapter / method	`siglip2`
Local reproduction code	`src/auto_research/reproductions/siglip2/`

Original paper¶

Motivation and method¶

SigLIP 2 extends SigLIP's pairwise sigmoid loss with captioning pretraining, global-local self-distillation, masked prediction, online data curation, and NaFlex dynamic resolution. This improves semantics, localization, dense features, and multilingual fairness.

flowchart LR
  I["Image encoder"] --> S["Sigmoid image-text objective"]
  T["Text encoder"] --> S
  I --> M["Masked prediction and self-distillation"]
  I --> C["Caption decoder"]
  S --> R["Shared SigLIP 2 representation"]
  M --> R
  C --> R

Key figure from the paper¶

Figure 3: Training flow and key optimization steps. From the original paper; copyright remains with the authors.

Core equation¶

\[ \mathcal L_{\mathrm{sigmoid}}= \frac1{|B|}\sum_{i,j}\log\left(1+\exp\left[-z_{ij} (t\,I_i^\top T_j+b)\right]\right),\quad z_{ij}\in\{-1,+1\}. \]

Paper results¶

All four model sizes outperform their SigLIP counterparts across zero-shot classification, retrieval, transfer, localization, and dense prediction. The authors release B/L/So400m/g checkpoints. No industrial online A/B is reported.

Local reproduction¶

Local comparison: Uniform label retrieval is 10.0%. Sigmoid contrastive learning plus masked-view self-distillation reaches 66.9% test accuracy (+56.9 points); shuffling images drops it to 10.8%.

auto-research reproduce --paper siglip2 --dataset-dir data --seed 42

See the fixed metrics.

Reproduction boundaries¶

Fashion-MNIST class text substitutes for multilingual captions. Sigmoid pair loss and center-masked student–teacher consistency train locally. Caption decoder, active curation, NaFlex, and localization head were not reproduced.