Task-Agnostic Compatible Adapter (TaCA)

Upgrade visual foundation models without retraining downstream tasks.

TaCA is a parameter-efficient adapter that enables seamless upgrades between visual foundation models (e.g. across CLIP variants) without retraining downstream tasks.

  • Learns a lightweight alignment between old and new backbones.
  • Preserves the downstream heads, so existing retrieval / recognition systems keep working during backbone upgrades.
  • Validated on large-scale video–text retrieval and video recognition benchmarks.

Paper