The authors define multi-turn contextual composed image retrieval, build a TIE model that encodes full dialogue history into evolving embeddings, and report state-of-the-art single-turn results plus large margins on their own 4-turn benchmark.
In: ICLR (2025)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
CoCo-IR: Contextual Composed Image Retrieval
The authors define multi-turn contextual composed image retrieval, build a TIE model that encodes full dialogue history into evolving embeddings, and report state-of-the-art single-turn results plus large margins on their own 4-turn benchmark.