PhysInsert inserts objects with geometry and scene interaction
Original title: PhysInsert: Physically Plausible Object Insertion with Geometric Priors and Bidirectional Object-Scene Interactions
The authors built Insert World and PhysInsert-Bench and report better fidelity plus a user study on placement and lighting.
PhysInsert is a reference-based object-insertion framework that jointly adapts an object and the surrounding scene for pose, illumination, and spatial consistency. The authors built Insert World from aligned Blender-rendered counterfactual pairs and object references across viewpoints and lighting. Geometry-Guided Adaptation combines a VGGT geometry stream with image and text streams, and 2D supervision covers object occupancy and environmental change. On PhysInsert-Bench and DreamBooth they report higher image fidelity; a user study found more accurate placement and better lighting, perspective, and pose.
Unconfirmed: Fidelity and user-study gains are reported by the authors only.
- Insert World uses strictly aligned Blender-rendered counterfactual pairs.
- PhysInsert-Bench adds cross-environment and cross-view cases.
- Geometry stream is VGGT under target-scene feature supervision.
- User study covers placement, lighting, perspective, and pose.
From the original article
1 min read at huggingface.co
We present PhysInsert, a framework for physically plausible reference-based object insertion that jointly adapts objects and their surrounding scenes. Existing methods often rely on masked inputs, generation-based pairs, or training pairs with limited viewpoint variation, leaving pose and illumination adaptation and bidirectional object-scene interactions insufficiently constrained.
By Yue Gong, Le Zhang, Rui Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.