acceptodds
Under review as a conference paper at ICLR 2027

Compositional Diffusion with Maximal-Noise Transitions for Long-Horizon Planning

Abstract

Compositional diffusion planners generate long-horizon trajectories by composing short local plans that are denoised jointly and kept consistent on their overlaps. However, when local plan distributions are multimodal, averaging incompatible local modes can produce infeasible composed plans. We attribute the persistence of mode-averaging to the predicted noise direction retained in the sampling transition. To address this challenge, we propose *Maximal-Noise Compositional Diffusion* (MaxCD), a *training-free* method that modifies sampling transitions on the overlaps. MaxCD constructs the next noisy state from the denoised estimate and fresh Gaussian noise, removing the predicted noise direction without additional denoiser evaluations. We additionally introduce a spectral split that applies maximal-noise transitions to low-frequency components and retains the original transition on high-frequency components to preserve fine details. Experiments demonstrate that MaxCD substantially improves long-horizon planning performance on OGBench and the visual consistency of compositional panorama generation. Project website at https://max-comp-diffusion.github.io/.

Primary area:
reinforcement learning
Keywords:
compositional diffusion, diffusion planning, long-horizon planning, trajectory stitching, offline reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_urLwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents MaxCD, a training-free method for improving consistency across overlapping segments in compositional diffusion planning. MaxCD heuristically modifies the denoising process to prevent samples from becoming trapped between incompatible modes. To preserve fine-grained details, it applies these modifications only to low-frequency components. Experiments across multiple domains demonstrate that MaxCD improves planning performance and reduces the occurrence of infeasible plans.

Critical strengths
  1. Experiments across multiple domains show that MaxCD significantly improves planning performance by reducing the occurrence of infeasible plans.
  2. As a training-free method, MaxCD can be readily integrated into a range of diffusion-based planning approaches.
  3. The method introduces little additional computational overhead.
Critical weaknesses

Maximal-noise transitions alter the sampling distribution even in the case of a single plan with an exact denoiser. The limitations section does not adequately discuss this distributional shift or its potential implications for consistency across overlapping plan segments.

Questions for authors

Could the distributional shift introduced by MaxCD negatively affect planning performance or consistency across overlapping plan segments? Under what conditions might such effects arise?

Reviewer_DwZYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes MaxCD, which changes the sampling transition of compositional diffusion planners. It adds fresh noise to low frequency components of overlapping segments while retaining the original transition elsewhere. Experiments cover OGBench planning and panorama generation.

Critical strengths
  1. The method is simple and requires no retraining or additional denoiser evaluations. The improvements on the Giant mazes are clear using the same pretrained planners and controllers.
  2. The toy example and analysis provide a clear motivation for changing the sampling transition to address persistent mode averaging.
  3. The evaluation covers locomotion, object manipulation, and panorama generation, using the same noise threshold and spectral bandwidth across tasks.
Critical weaknesses
  • The offline RL comparisons use different execution budgets. Appendix G imports OGBench scores, whereas Table 5 increases several episode limits.
Questions for authors
  • Proposition 1 establishes escape from an idealized midpoint region, rather than convergence to an executable plan. Figure 8 also shows heterogeneous geometric validity: MaxCD averages 86%, but achieves 60% on Task 3 versus CDGS’s 75%. Relating selected-plan validity to tracking failures and replanning would clarify which failures the method resolves.
Reviewer_xQQtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper attributes persistent mode-averaging in compositional diffusion planning to the predicted noise direction retained in the DDIM/DDPM transition. MaxCD replaces this transition on the overlaps of interior local plans with a maximal-noise transition, σ_t = sqrt(1 − α_{t−1}), which gives subsequent denoising a chance to pull the neighboring local plans toward a common mode. It is applied only to low-frequency components and only when t/T > γ. It needs no training and no extra denoiser evaluations. On OGBench, it improves CompDiffuser on the Giant stitch mazes, and it is also applied to panorama generation.

Critical strengths
  • The paper is clear and easy to follow.
  • The method is training-free and adds only 7–10% planning time over CompDiffuser (Table 10).
  • The gains over CompDiffuser on the Giant mazes are large.
Critical weaknesses
  1. The novelty is limited: the maximal-noise transition (Eq. 10) is the maximum stochasticity already used by StochSync (arXiv:2501.15445), which is not cited, so the new method components are the overlap mask, the spectral split, and the threshold γ.
  2. The closest training-free baseline, RCD (arXiv:2605.03075), is not compared, although it reports higher success on PointMaze-Giant and AntMaze-Giant stitch (100% and 89% versus 94% and 88%) with the same pretrained CompDiffuser models.
  3. No experiment varies the noise level between the planning base sampler (η = 1, Table 6) and the maximum, so it is unclear whether the maximal value, rather than any extra noise, drives the gains.
  4. Proposition 1 only bounds the probability of staying in the mode-averaged region under a fixed-midpoint model. It does not show that MaxCD reaches a consistent mode, and 14% of MaxCD plans on AntMaze-Giant-Stitch still pass through walls (Section 4.1).
Questions for authors
  1. I suggest that the second paragraph of the introduction clarify which limitation of CDGS, ECD, and SyncDiffusion MaxCD overcomes, rather than listing them.
  2. I will raise my score if MaxCD keeps a clear margin over RCD and over an intermediate-noise control, and if the novelty over StochSync is clarified.

Comments on the paper

1 comment

Dear reviewers,

The authors have now submitted their responses. Please review them and share your thoughts on whether your concerns have been addressed. I encourage you to engage with one another’s assessments, discuss any remaining disagreements, and update your reviews and scores as appropriate.

Thank you.

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.