Compositional Visual Planning via
Inference-Time Diffusion Scaling

Yixin Zhang1, Yunhao Luo1,2, Utkarsh A. Mishra1, Woo Chul Shin1, Yongxin Chen1, Danfei Xu1
1Georgia Institute of Technology 2University of Michigan
ICLR 2026

Evaluation on Compositional Planning Benchmark: Puzzle Scene

All evaluation episodes of the Puzzle scene: start and goal images, synthesized video plans and policy rollouts, for our method and DiffCollage. ← Back to the project page

For pixel-based evaluations (synthesized videos), we disable ray tracing to ensure fully observable scenes without shadows or lighting artifacts. For policy rollouts, ray tracing is enabled to provide physically realistic observations and a clearer visualization of the robot’s manipulation skills.