DRIFT: Residual Flow Adapter for VLM Continuous Outputs

Arxiv CS.CV (computer vision)

Proposes DRIFT, a residual flow adapter that decodes continuous outputs in vision-language models by modeling residual prediction flows. Improves visual grounding and referring segmentation tasks, addressing limitations of discrete token decoding.

vision-language-models

Arxiv CS.CV (computer vision)·Zhuoming LiuJinhong LinKwan Man ChengLin ZhangShayok BagchiYin Li

6 days ago