Zhejiang Univ Open-Sources PhyEdit for 3D-Aware Image Editing

量子位 · wechat · 2026-08-13

The ReLER team at Zhejiang University has proposed and open-sourced PhyEdit, a DiT-based image editing model guided by explicit 3D geometric previews. It solves common failures in traditional models when handling object scale, occlusion, and multi-object manipulation.

The method decouples geometry from generation: it estimates scene depth and camera parameters, moves object point clouds in 3D space to create a rough preview, and then uses a generative model to recover textures and details. The team also released the RealManip-40K dataset and the ManipEval benchmark. Experiments show PhyEdit significantly outperforms existing commercial models in 3D accuracy while retaining standard editing capabilities.

Original post →

More from Multimodal

Multimodal channel →