TARA adds token-focused masking and a token alignment loss to LoRA adapters, allowing several independently trained personalized adapters to be composed with less identity loss and feature leakage.
Emerg- ing properties in self-supervised vision transformers
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
TARA: Token-Aware LoRA for Composable Personalization in Diffusion Models
TARA adds token-focused masking and a token alignment loss to LoRA adapters, allowing several independently trained personalized adapters to be composed with less identity loss and feature leakage.