CVPR 2026 | CorrAdapter：让多图扩散模型先“对齐”，再生成

TL;DR - CVPR 2026 paper CorrAdapter improves cross-view and temporal consistency in multi-image diffusion models by aligning internal features before generation. Its plug-and-play design works without external depth, segmentation, optical flow, or explicit geometry.

• Extracts “diffusion-native” correspondences from cross-image query-key similarities in intermediate features.
• Aggregates information only within locally matched regions, reducing structural and texture drift without indiscriminate feature mixing.
• Improves multi-view and text-to-video consistency across several baselines, including SyncDreamer, MVAdapter, Wan2.1, and HunyuanVideo.
• The training-free version adds inference-time and memory overhead; video stability gains can also reduce motion magnitude.