acceptodds
Withdrawn from ICLR 2027

Diffusive Dynamics Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusive Dynamics Modeling uses time as the control variable to predict the diffusion of substances at arbitrary future time points based on current images, which provides a foundation for future generation and real-world modeling. However, this task remains challenging due to: (1) Time-controlled generation of dynamics lacks temporal connections between adjacent time points, leading to discontinuities. (2) Prediction deviations across temporal spans violate the macroscopic directionality, leading to inconsistency. The diffusion model generation process follows diffusive dynamics along the noise-level axis, while real diffusive dynamics evolve along the time axis. Taking these two dynamics as the two axes to form a two-dimensional grid enables the establishment of temporal relationships across multiple noise levels. Therefore, we propose Diffusive Grid, a unified diffusive dynamics modeling framework to establish temporal relationships throughout the denoising process. Based on this grid, Dynamics Alignment Learning (DAL) aligns predictions that reach the same grid node through temporal evolution and denoising, thereby constraining the entire generation process with adjacent time points and improving continuity. Moreover, Grid Walk (GW) constructs paths across temporal spans and anchors their endpoints to the target, allowing supervision to propagate through intermediate states and regularize consistency. Experiments covering both natural and physiological diffusive processes demonstrate that our method achieves better overall performance in continuity and consistency.

Primary area:
generative models
Keywords:
Diffusion model, diffusive dynamic

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_trNtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies time-conditioned prediction of diffusive processes: given a current image and an arbitrary target time, the goal is to generate the future observation. It proposes Diffusive Grid, which treats the denoising axis of a diffusion model and the physical time axis as a two-dimensional grid of noisy latent states. In addition to the conditional Denoiser, an Evolver network predicts transitions along time at a fixed noise level. Dynamics Alignment Learning (DAL) aligns the Evolver's prediction of a grid node with the Denoiser's prediction of the same node from a higher noise level, and Grid Walk (GW) supervises the endpoints of multiple paths that alternate between temporal evolution and DDIM denoising. On Scalarflow and Duke DCE-MRI, the method reports better frame-level fidelity and lower temporal (tRMSE) and drift (dRMSE) errors than DDPM, OU-noise, STCL, and Dynamical Diffusion.

Critical strengths
  1. Discontinuity and directional inconsistency are real failure modes of generators that produce each target time independently, and the paper separates them clearly with dedicated metrics.
  2. Coupling a learned temporal transition model with the denoising process at multiple noise levels is an intuitive idea, and the method retains direct target-time generation at inference.
Critical weaknesses
  1. The connection to physical diffusion is largely nominal, and the novelty relative to existing time–noise modeling is unclear. The grid is motivated by an analogy between the "diffusive dynamics" of denoising and real diffusive processes, but the forward noising process of a diffusion model is unrelated to the physical transport being modeled, and the method uses no physical structure such as a diffusion equation, mass conservation, or an explicit directionality constraint. The method is therefore best viewed as a generic regularizer for time-conditioned image prediction. From this perspective, existing methods such as Diffusion Forcing, Rolling Diffusion, FIFO-Diffusion, and AR-Diffusion already vary noise levels or denoising schedules across sequence positions. DYffusion also connects its generative steps to the physical time of the underlying dynamics.

  2. The mechanism by which the training-time grid improves inference is not established. At inference, the model generates each target time separately using only the Denoiser. The paper should clarify whether the same initial Gaussian noise is reused across target times and whether this protocol is consistent across methods, since noise coupling can affect tRMSE and dRMSE. The straight path in GW uses only the Evolver at noise level 0 and cannot directly affect the inference-time Denoiser, consistent with this, the Straight Path row in Table 5 is identical to the +DAL row in Table 1, which contradicts the claim that the two path types provide complementary benefits. Finally, Table 4 shows that rolling out the Evolver on the initial noise substantially improves results on DCE-MRI, but this procedure is not used at inference and the reason is not discussed.

Questions for authors
  1. Is the initial noise shared across target times at inference, and is the same protocol used for all methods?
  2. Why is the Straight Path result in Table 5 identical to the +DAL result in Table 1, given that the straight path does not involve the Denoiser?
  3. Why is the Evolver rollout in Table 4 not used at inference, given its large improvements?
  4. Can the authors compare with Diffusion Forcing and/or DYffusion, and clarify which parts of the method are specific to diffusive processes?
Reviewer_tXimOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces diffusive dynamics modeling addresses prediction of future images from an initial observation and a target time. The authors propose Diffusive Grid to align the diffusive dynamics throughout the denoising process. Based on this grid, the authors introduce two training objectives that connect prediction with denoising. First, **Dynamics Alignment Learning (DAL) ** aligns predictions from a temporal Evolver and a diffusion Denoiser. Second, ** Grid Walk ** constructs the paths across paths that combine temporal evolution and denoising. The experimental setup uses ScalarFlow smoke sequences and breast DCE-MRI. The reported results show improved fidelity and temporal coherence over standard DDPM.

Critical strengths
  • A coherent methodological contribution: The diffusive grid connects physical-time and denoising explicitly, as DAL and GW provide complimentary supervision. The Evolver guides training, while the Denoiser directly predicts frames at requests time-indices without requiring recursive temporal prediction. Furthemore, this leads to the main practical appeal of this method is that the inference cost stays equal to the base model avoiding the error accumulation of the recursive Evolver.
Critical weaknesses
  • Baseline comparisons provide limited evidence of superiority: The component ablations and straight-versus-detour comparison are informative, but do not establish how much the gains depend on alignment across noise levels rather than simpler temporal supervision or additional training computation. A clean-state-only alignment control and a GW-only variant under matched training budgets would clarify this contribution. Some improvements over competing methods are also small, particularly DCE-MRI tRMSE, whereas others are substantial. Since the reported significance tests compare successive ablation variants, the statistical reliability of improvements over the strongest competing baselines remains unclear.
Questions for authors
  1. What sampler and number of denoising steps are used at inference, and how are Gaussian initializations—and any stochastic denoising updates—coupled across target times? Is this protocol identical across DDPM and the DAL/GW variants? How sensitive are the temporal metrics and relative improvements to shared versus independent noise?

  2. Could you compare alignment across multiple noise levels with a clean-state-only alignment control under comparable training budgets? Does the λDAL=0\lambda_{\mathrm{DAL}}=0 setting in Fig. 5(b) retain GW throughout training, and could you provide its numerical results? What additional training cost does the full method introduce?

  3. Could you report paired comparisons or uncertainty estimates against the strongest competing baselines, particularly for the small DCE-MRI tRMSE improvement over OU-noise? Since individual axial slices are treated as samples, please clarify whether evaluation and significance testing aggregate results by patient or otherwise account for dependence among slices from the same patient. Please also clarify how DyDiff was adapted and tuned for single-frame conditioning.

1 comment

Thank you for the thorough response and the additional experiments. The clarifications on the inference protocol, the patient-level evaluation with paired tests, the training cost and the DyDiff adaptation address my questions on those points.

However, I am inclined to keep my score. I find that the new results narrow the paper's central claims more than they support them. The clean-state-only control recovers most of the gain (SSIM 0.817 against 0.828 and tRMSE 0.155 against 0.150 for full DAL, within the reported standard deviations), so the benefit of aligning across noise levels remains weakly supported. Also under matched compute, DAL alone outperforms DAL+GW on four of five metrics, while GW raises the per-step cost. Finally, DYffusion and Diffusion Forcing obtain lower tRMSE on both datasets and lower dRMSE on DCE-MRI, which does not support the claim of better continuity and consistency.

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

The paper proposes Diffusive Grid(DG), a unified framework for spatial-temporal modeling, which attempts to establish temporal relationships throughout the denoising process. DG first trains a evolver to learn from temporal relationship of noisy latents. After that, DG introduces dynamics alignment learning at align the denoiser and the evolver. Additionally, grid walk is proposed, which sample the arbitory path to leverage the knowledge of evolver to prompt the temporal comsisitency of denoiser. Experiments on Scalarflow and DCE-MRI shows its effectiveness.

Critical strengths
  1. **Straightforward Inference Pipeline. **Although Diffusive Grid introduces multiple training stages and loss functions, at the inference time, DG directly produces the target forecasts conditioned on the initial frame and target time and not requiring the generation of intermediate frames. The evolver and Grid Walk are not needed during inference, keeping the inference pipeline straightforward.
  2. **Exploration with Grid Walk. **The design concept behind Grid Walk is appealing: By randomly exploring paths along the denoising and prediction-time axes, Grid Walk improves consistency across frames and enables model to capture low-frequency signals at high noise levels.
Critical weaknesses
  1. **Comparison with Diffusion Forcing. ** Diffusion Forcing [1] introduces a spatial-temporal diffusion framework that trains on video or time-series sequences with independently sampled per-frame noise levels, which also provides an arbitrary sampling space to enhance temporal connections and frame consistency. DG shares a similar design philosophy with Diffusion Forcing, and it will be helpful to include Diffusion Forcing as a baseline to clarify the benefits of DG’s training strategy.

  2. Limited Experiment Scope. Experiments are limited to Scalarflow and DCE-MRI, both of which consist of grayscale image sequences. Evaluation on more diverse datasets, including natural RGB videos, would help assess the generalization of DG. Moreover, some spatiotemporal tasks involve video-to-video prediction, requiring the generation of a coherent future sequence conditioned on observed frames. DG’s applicability to this setting remains unclear.

Reference

[1] Boyuan Chen, Diego Martí Monsó, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion. Advances in Neural Information Processing Systems, 37, 2024.

Questions for authors
  1. Could the authors provide an analysis of additional training cost introduced by DAL and Grid Walk? Do the improvements persist when the baselines are given a same compute budget?
  2. Is DG implemented as a latent diffusion model? If so, which encoder–decoder is used, and is it pretrained and frozen during training?
  3. How do the authors think about the trade-offs between full-sequence prediction and direct target-time prediction like DG? I would appreciate the authors’ perspective on this question without additional experiment.

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

Related papers

Discussion (0)

Sign in to comment.