A frozen LLM text encoder (NV-Embed-V2) can guide image encoder training and outperform vanilla CLIP on compositional benchmarks and long captions.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Language-Image Alignment with Fixed Text Encoders
A frozen LLM text encoder (NV-Embed-V2) can guide image encoder training and outperform vanilla CLIP on compositional benchmarks and long captions.