Diffusion models excel at short-horizon robot planning, yet scaling them to long-horizon tasks remains challenging due to computational constraints and limited training data. Existing compositional approaches stitch together short segments by separately denoising each component and averaging overlapping regions. However, this suffers from instability as the factorization assumption breaks down in noisy data space, leading to inconsistent global plans. We propose that the key to stable compositional generation lies in enforcing boundary agreement on the estimated clean data (Tweedie estimates) rather than on noisy intermediate states. Our method formulates long-horizon planning as inference over a chain-structured factor graph of overlapping video chunks, where pretrained short-horizon video diffusion models provide local priors. At inference time, we enforce boundary agreement through a novel combination of synchronous and asynchronous message passing that operates on Tweedie estimates, producing globally consistent guidance without requiring additional training. Our training-free framework demonstrates significant improvements over existing baselines, effectively generalizing to unseen start-goal combinations that were not present in the original training data. Project website: https://comp-visual-planning.github.io/
Figures & tables
Figure 1: Compositional Visual Planning via Inference Time Diffuser Scaling. We train a short-horizon visual diffusion model on clips treated as a single factor. At inference, we scale visual planning horizon without retraining by chaining overlapping factors into a linear factor graph: the start and goal boundary variables are anchored at the ends, while neighboring factors exchange information through shared transition boundary variables.
Figure 2: Motivating toy example. We train a short-horizon diffusion model on circular arc clips (left). At test time, three 120∘ arc generators are composed to form a three-petal “flower”.
Figure 3: Real World Task Layout. This task involves 2 start and 2 goal configurations. We also evaluate on more challenging tasks. A complete list of task settings is provided in Appendix E .
Figure 4
Figure 6: Hardware Setup. We deploy our method on a Franka Emika Panda robot.
Real Scene
IND:Task1
IND:Task2
OOD:Task3
OOD:Task4
DiffCollage
1/10
1/10
0/10
0/10
Ours
9/10
7/10
10/10
8/10
Table 3: Real-robot success rates. Our method substantially outperforms DiffCollage across both in-distribution (IND) and out-of-distribution (OOD) tasks on real hardware.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Visualization of Tool Scene.
Figure 8: Visualization of Drawer Scene.
Figure 9: Visualization of Cube Scene.
Figure 10: Visualization of Puzzle Scene.
Hyperparameter
Value
Diffusion Time Step
500
Batch Size
512
Optimizer
Adam
Learning Rate
1×10−4
Iterations
1M
Discount Factor γ
0.6
Appendix
Table 4: Relevant hyperparameters used in our experiments.
Scene
# Models Composed
Sampling Time (s) ↓
DiffCollage
Ours
Tool-Use
3
7.1
17.2
Drawer
5
18.9
30.4
Cube
5
21.1
31.7
Puzzle
6
30.4
61.8
Appendix
Table 5: Sampling time during deployment. This is measured as the mean wall-clock time across all samples within a single scene.
Figure 11: Task1 Synthetic Plan.
Figure 12: Task2 Synthetic Plan.
Figure 13: Task3 Synthetic Plan.
Figure 14: Task4 Synthetic Plan.
Figure 15: Failure Mode of Synchronous Message Passing: Case 1.
Figure 16: Failure Mode of Synchronous Message Passing: Case 2.
Diffusion-based planning generates diverse action sequences or trajectories by progressively denoising candidates conditioned on goals, scenes, and constraints.
Diffusion models provide strong priors for generating structured data, but many tasks require outputs beyond the scale on which these models are typically trained. Compositional generation addresses this by composing overlapping local plans from a pretrained short-horizon prior into a long-horizon output. However, standard composition primarily enforces agreement between neighboring local plans, yielding local consistency without directly specifying the global structure of the full composition. As a result, locally compatible plans may still form an implausible route, task sequence, or temporal evolution. Existing methods improve global coherence by repeatedly propagating local consistency signals or by adding inference-time optimization, but these procedures become expensive as the number or dimensionality of local plans increases. We propose Coarse-to-Fine Compositional Diffusion (CoFi), an inference-time sampler that separates global structure formation from local detail refinement. CoFi first aligns local denoised estimates around a shared coarse structure, producing a global scaffold that captures the long-range task-level arrangement. It then diffuses this scaffold to an intermediate noise level and denoises it with the same pretrained local prior, restoring local fine structure while preserving the scaffold-induced global coherence. Across long-horizon robotic planning, panoramic image generation, and long video generation, CoFi not only improves both global coherence and local sample quality over prior compositional baselines, but also requires 2-8x fewer denoiser evaluations.
Compositional diffusion planning generates long-horizon trajectories by stitching together overlapping short-horizon segments through score composition. However, when local plan distributions are multimodal, existing compositional methods suffer from mode-averaging, where averaging incompatible local modes leads to plans that are neither locally feasible nor globally coherent. We propose Refining Compositional Diffusion (RCD), a training-free guidance method that steers compositional sampling toward high-density, globally coherent plans. RCD leverages the self-reconstruction error of a pretrained diffusion model as a proxy for the log-density of composed plans, combined with an overlap consistency term that enforces consistency at segment boundaries. We show that the combined guidance concentrates sampling on high-density plans that mitigate mode-averaging. Experiments on challenging long-horizon tasks from OGBench, including locomotion, object manipulation, and pixel-based observations, demonstrate that RCD consistently outperforms existing methods.
Kyowoon Lee, Yunhao Luo, Anh Tong +1
KAIST · University of Michigan · Korea University +1
Compositional diffusion models offer a promising route to long-horizon planning by denoising multiple overlapping sub-trajectories while ensuring that together they constitute a global solution. However, enforcing local behavior over long chains is often insufficient for a coherent global structure to emerge. Recent works tackle this limitation through intrinsic search, which explores multiple paths during the denoising process. While intrinsic search improves global coherence, it comes at the cost of repeated evaluations of an already compute-heavy model. In this work, we argue that extrinsic search, performed outside the denoising process, offers a more effective mode of exploration for long-horizon planning while naturally enabling the use of classical algorithms to solve unseen combinatorial tasks at test time. Our eXtrinsic search-guided Diffuser (XDiffuser) first computes a plan over a state-space graph -- serving as a lightweight local connectivity oracle for the diffusion model. The plan is then used to guide denoising for a single trajectory, effectively offloading the burden of exploration. XDiffuser outperforms diffusion-based baselines on long-horizon tasks, with particularly large gains in the low-quality data regime and on unseen tasks beyond goal-reaching, including multi-agent coordination and TSP-style reasoning. Project website: https://yanivhass.github.io/XDiffuser-site/