ReVision adapts frozen image generators to perception tasks with per-instance parameter updates

The framework maps image-instruction pairs to transient low-rank updates, yielding competitive segmentation, depth, and surface-normal prediction without task-specific heads.

Chinese Tech
Yitong Yang · Yinglin Wang · Xuexin Liu · Wei Chen · Zhaoyang Xu · Jiaxin Lin · +3 more

Tencent Hy · Shanghai University of Finance and Economics

Research Digest··2 min read
The authors introduce ReVision, which connects a frozen multimodal understanding model to a frozen image generator through a trainable parameter generator.

ReVision takes an image and a free-form instruction as input.

§