3D human object interaction synthesis results by InteractAnything. Our method enables the generation of diverse, detailed, and novel interactions for open-set 3D objects. Given a simple text description with goal interaction and any object mesh as input, we can synthesize different natural HOI results without training on 3D assets. The orange and green boxes of (b) indicate detailed contact poses from different views.

Framework of InteractAnything. Given a text description and an open-set object mesh as input, our approach begins by querying LLM feedback to infer precise human-object relationships, which are used to initialize object properties (Section 3.2). Next, we analyze the contact affordance on the object geometry (Section 3.3). The human pose is synthesized using a pre-trained 2D diffusion model, guided by SSDS loss and designed spatial constraint (Section 3.4). Finally, based on the targeted object contact areas and a plausible human pose, we perform expressive HOI optimization to synthesize realistic and contact-accurate 3D human object interactions (Section 3.5).