Researchers have developed VIBRA, a novel method for backward-compatible cross-modal retrieval. This technique allows new encoders to query galleries encoded by older models, eliminating the need to re-encode vast datasets when upgrading systems. VIBRA utilizes a consensus barycenter to correct single-model errors and trains new encoders to inherit this consensus, ensuring that even absent models influence query relevance. Tested across five benchmarks and six different CLIP, MetaCLIP, and SigLIP2 backbones, VIBRA outperformed existing compatible methods in all evaluated settings, significantly improving Recall@1 while adding minimal latency. AI
IMPACT Enables more efficient and cost-effective upgrades of large-scale cross-modal retrieval systems.
RANK_REASON The cluster describes a new research paper detailing a novel method for cross-modal retrieval.
Read on arXiv cs.IR (Information Retrieval) →
- arXiv
- Hugging Face
- SigLIP2
- VIBRA
- CatalyzeX
- Connected Papers
- CORE Recommender
- DagsHub
- Gotit.pub
- Litmaps
- ScienceCast
- scite Smart Citations
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →