Research map / Image generation / Fast diffusion sampling
Compositional Visual Planning via Inference-Time Diffusion Scaling
Yixin Zhang, Yunhao Luo, Utkarsh Aashu Mishra, Woo Chul Shin, Yongxin Chen, Danfei Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enforce boundary agreement on Tweedie estimates instead of noisy-state score averaging (vs DiffCollage) | compositional video planner · Compositional Planning Bench Overall (OOD) | success rate (%): 0 -> 54 | Improved |
| use the training-free method versus trained policies on unseen start-goal pairs | compositional video planner vs GCDP · Compositional Planning Bench Overall (OOD) | success rate (%): 15 -> 54 | Improved |
| add asynchronous loss to synchronous message passing (ablation) | message passing sampler · Cube scene | success rate | Improved |
| add synchronous loss to asynchronous message passing (ablation) | message passing sampler · Cube scene | success rate | Improved |
| increase number of diffusion sampling steps (ablation) | compositional video planner · Drawer scene | success rate | Improved |
| remove discount factor in message passing (ablation) | synchronous message passing · qualitative failure cases | start/goal spatial alignment | Hurt |
| replace DiffCollage with the proposed sampler in real-world tasks | real robot Franka · Real robot Task3 (OOD) | success (out of 10): 0/10 -> 10/10 | Improved |
| use proposed method rather than DiffCollage for visual quality | video plans · VBench++ Aesthetic (OOD) | aesthetic quality: 0.46 -> 0.48 | No clear effect |
| use proposed method rather than DiffCollage | video plans · VBench++ Imaging Quality (OOD) | imaging quality: 0.55 -> 0.69 | Improved |
| use proposed method rather than DiffCollage (test-time backprop) | sampling · Puzzle sampling time | sampling time (s): 30.4 -> 61.8 | Hurt |
| use proposed method rather than best policy baseline on in-distribution tasks | Drawer scene · Compositional Planning Bench Drawer (IND) | success rate (%): 54 -> 53 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiffCollage — DiffCollage: Parallel Generation of Large Content With Diffusion Models (CVPR 2023) · better on 7 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning (NeurIPS 2026 · similarity 0.95)
- Generative Trajectory Stitching through Diffusion Composition (NeurIPS 2025 · similarity 0.88)
- LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE (NeurIPS 2026 · similarity 0.88)
- Compositional Monte Carlo Tree Diffusion for Extendable Planning (NeurIPS 2025 · similarity 0.88)
- Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts (CVPR 2024 · similarity 0.87)
- Amortizing intractable inference in diffusion models for vision, language, and control (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
