- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper examines diffusion training from a noise-sampling perspective, arguing that appropriate noise schedules vary across training stages. It proposes a PPO-based adaptive scheduler and demonstrates its effectiveness on standard benchmarks. Overall, the work offers a compelling perspective, solid theoretical grounding, and strong empirical results.
Critical strengths
Strengths
- The paper offers an interesting perspective by treating forward noise scheduling as an adaptive training decision rather than a fixed design choice.
- The proposed schedule is compact and interpretable: two parameters control its shape, while fixed endpoints and monotonicity preserve a valid noising process.
- The PPO-based controller is paired with a structured reward design, and the ablation studies provide evidence for the contributions of its main components.
- The evaluation covers multiple datasets and random seeds under matched backbones and sampling protocols. The reported early-stage results and training overhead also help assess the performance–cost trade-off.
- The manuscript is well organized and clearly written.
Critical weaknesses
Weaknesses and questions
- The PPO reward components appear largely heuristic and lack sufficient theoretical justification, which may limit their interpretability and transferability.
- The evaluation is confined to U-Net backbones. Although this enables controlled comparisons, experiments with more recent transformer-based architectures would strengthen the generality of the results.
- Most baselines use Gaussian noising–denoising diffusion. It would be valuable to test whether adaptive time scheduling provides similar benefits in flow-matching models.
- The interpretation of Figure 4 needs clarification. The schedule is selected before the current batch is sampled, while the policy observes only lagged batch features. How, then, does the method systematically assign higher noise to harder current batches? The authors should clarify how batch difficulty is measured and what produces the reported correlation.
Questions for authors
See the weakness
We thank the reviewer for the positive assessment of the perspective and the empirical basis, and for the question about the timing of the batch feature in Figure 4. To make the data flow more explicit, we have updated the caption to clarify that fk is computed from the clean minibatch before the action and before noise is added. We have also added controls to separate the current-minibatch signal from adjacent batches. Figure 4 now clearly reflects a current-minibatch correlation.
Furthermore, the revised manuscript clarifies the system mechanics: a shared (m,n) covers the entire minibatch, L0 functions strictly as a held-out difficulty label, and the FID improvement over cosine is diven by the phase-dependent clipped schedule. The following sections detail the reward terms, backbone scope, flow-matching extension, and Figure 4 timing mechanics.
1. Reward terms and ablation
In response to the question on heuristic reward components, we clarify the theoretical grounding of the alignment term and report the sensitivity of the core rewards.
Ralign is the inner product of the binned clipped weight and the residual EMA. With fixed endpoints, the unclipped objective is invariant to (m,n); after clipping, the bin masses still depend on (m,n), and this inner product is the part of the schedule that the reward can still adjust. Rcyc measures preview-denoiser recoverability. Rtemp checks whether the residual decreases after a short horizon H. Rlearn penalizes distance from a target residual of 0.4. Rcost and Rsmooth regularize noise magnitude and action jumps.
At 200 epochs on CIFAR-10, setting λcyc to 0 gives FID 8.89; setting λalign to 0 gives 8.28; the default weights give 7.19. Among these, Rcyc and Ralign show the larger FID sensitivity.
2. U-Net backbone
In response to the concern regarding the U-Net backbone, we clarify that our evaluation is intentionally scoped to this matched setting to ensure controlled comparisons.
All image entries in the main result table use the same U-Net backbone and 1000-step ancestral sampling (about 35.7M parameters at 32×32; about 113M at 256×256).
3. Flow matching
In response to the question about generalization to flow-matching models, we report a separate 200-epoch rectified-flow experiment. We clarify the matched baseline to isolate the mechanism transfer.
NFDM is the neural-flow baseline trained on the same U-Net. The same sigmoid schedule maps uniform time to interpolation time, the reward keeps Rcyc, Ralign, and Rsmooth, and the loss is the velocity loss. Uniform time reaches FID 10.52±0.22; the time warp reaches 9.61±0.14 (revised-appendix flow panel).
This experiment is designed to test the transferability of the time-warping mechanism under a strictly matched 200-epoch velocity-loss budget. The relevant baseline is therefore the uniform-time run, which our method improves upon by a clear margin. The full-training NFDM Gaussian FID of 7.65 is reported elsewhere in the main table as a separate baseline.
4. Figure 4 timing
In response to the question about the batch feature timing, we clarify the exact data flow before noise addition. We show that the reported correlation acts on the current minibatch without any leakage.
The clean minibatch is drawn first. The batch feature fk is a frozen ResNet-18 feature of those clean images, pooled over the minibatch and computed before the policy samples (m,n) and before noise is added. This feature does not depend on the denoiser loss of the current step.
Batch difficulty L0 is the MSE of a separate frozen reference denoiser at a fixed noise level. It is used only as a held-out difficulty label in Figure 4, not as an input to the policy. The correlations with the current minibatch are R2=0.07, 0.52, and 0.87 (early, mid, late). The previous minibatch gives 0.06, 0.31, and 0.47; adjacent minibatches give 0.21, 0.12, and 0.16. In the late stage, the current minibatch is clearly the stronger correlate, and the adjacent-batch correlation of 0.16 does not account for the late-stage R2 of 0.87.
One slope and one inflection cover the entire minibatch. In 63.1% of minibatches the within-batch spread of L0 exceeds the easy--hard gap. The correlation reported above operates at the minibatch level.