LatentDiff: Scaling Semantic Dataset Comparison to Millions of Images
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2026
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866917454043676672 |
|---|---|
| author | Flora, James Thopalli, Kowshik Kulkarni, Akshay R. Wong, Weng-Keen Liu, Shusen |
| author_facet | Flora, James Thopalli, Kowshik Kulkarni, Akshay R. Wong, Weng-Keen Liu, Shusen |
| contents | We present LatentDiff, a scalable framework for semantic dataset comparison that operates directly in the latent space of pretrained vision encoders. By combining sparse autoencoder-based divergence testing with density ratio estimation, LatentDiff identifies interpretable semantic differences between datasets at a fraction of the computational cost of caption-based alternatives. We also introduce Noisy-Diff, a benchmark capturing realistic sparse distribution shifts that cause existing methods to struggle. Experiments demonstrate that LatentDiff achieves superior accuracy while remaining robust to settings where an extremely small fraction of images (from 5% to <1% ) differ semantically. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_00899 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | LatentDiff: Scaling Semantic Dataset Comparison to Millions of Images Flora, James Thopalli, Kowshik Kulkarni, Akshay R. Wong, Weng-Keen Liu, Shusen Computer Vision and Pattern Recognition Machine Learning We present LatentDiff, a scalable framework for semantic dataset comparison that operates directly in the latent space of pretrained vision encoders. By combining sparse autoencoder-based divergence testing with density ratio estimation, LatentDiff identifies interpretable semantic differences between datasets at a fraction of the computational cost of caption-based alternatives. We also introduce Noisy-Diff, a benchmark capturing realistic sparse distribution shifts that cause existing methods to struggle. Experiments demonstrate that LatentDiff achieves superior accuracy while remaining robust to settings where an extremely small fraction of images (from 5% to <1% ) differ semantically. |
| title | LatentDiff: Scaling Semantic Dataset Comparison to Millions of Images |
| topic | Computer Vision and Pattern Recognition Machine Learning |
| url | https://arxiv.org/abs/2605.00899 |