acceptodds
Under review as a conference paper at ICLR 2027

Defect-Guided Time-Warp Flow Distillation for Semigroup-Robust Fast Diffusion Generation

Abstract

Fast diffusion generation has been improved through few-step distillation, and trajectory rectification. However, a central difficulty remains: a student model only trained on single-step transitions is deployed via repeated composition across different step counts. This resulting discrepancy between a direct transition and its multi-step composition is defined as the semigroup defect. We therefore recast fast diffusion distillation as a time-geometry adaptation problem, where temporal allocation along the trajectory is optimized instead of explicitly modifying the state-space path. Specifically, we propose Defect-Guided Time-Warp Flow Distillation (DG-TWFD), which jointly learns a continuous-time student flow map and a monotone time warp that reallocates temporal resolution toward teacher regions with high compositional difficulty. Training integrates teacher flow matching with defect-aware optimization and a time-warped curriculum. We evaluate our method on unconditional CIFAR-10 and ImageNet-64, achieving competitive Fréchet Inception Distance (FID) and Inception Score (IS) on CIFAR-10, and FID and recall on ImageNet-64 across different inference step settings. Our best results are 1.81 FID / 9.87 IS on CIFAR-10 and 2.49 FID / 0.67 recall on ImageNet-64. All code and data used in this work are available at https://anonymous.4open.science/r/DGTW-code-base-95F0/Anonymous Github link.

Primary area:
generative models
Keywords:
diffusion, generative models, image processing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8oNaOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Defect-Guided Time-Warp Flow Distillation (DG-TWFD), which focuses on trajectory-based diffusion distillation. DG-TWFD jointly learns a continuous-time student flow map and a monotone time warp that reallocates temporal resolution toward regions with high compositional difficulty. By explicitly minimizing the semigroup defect, which measures the discrepancy between a direct transition and its multi-step composition, DG-TWFD aims to improve the robustness of diffusion generation across different inference step counts. The authors evaluate their method on CIFAR-10 and ImageNet-64.

Critical strengths
  • The paper is well-written and easy to follow.
  • The paper presents an interesting attempt to introduce adaptive timestep scheduling based on the compositional difficulty observed during training.
Critical weaknesses
  • Limited evaluation on large-scale generation tasks. The experimental evaluation primarily focuses on CIFAR-10 and ImageNet-64. More comprehensive experiments on higher-resolution datasets, such as ImageNet-256, or text-to-image generation tasks would help demonstrate the scalability and generalizability of the proposed method.
  • Insufficient ablation of the time-warping mechanism. The proposed method combines defect-aware training-time sampling with an adaptive inference schedule. However, the individual contributions of these two components are not sufficiently disentangled, making it difficult to determine which component is primarily responsible for the performance improvements.
Questions for authors
  • What are the individual contributions of training-time sampling and inference-time scheduling? Could the authors provide an ablation study that separates these two components to clarify their respective contributions to the performance gains?
  • Can the authors provide comparisons under matched computational budgets? Since DG-TWFD introduces additional operations for semigroup defect estimation and adaptive time warping, it would be helpful to compare its performance with existing distillation methods under matched training computational budgets.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 8oNa

We thank the reviewer for the constructive comments. We answer each point below.

W1 (large-scale evaluation).

The ImageNet-256 experiment is in Appendix F.2 (latent space): the warp lowers the defect by 6–12% at every checkpoint, which shows that the method scales to 256×256256 \times 256. Higher resolutions and text-to-image use the same latent formulation and would not add further insight (see also our reply to TGbi, W3).

W2 (ablation of the two components).

Both components contribute, and they are coupled by design: one set of bin masses defines the training sampling and the inference grid (Eqs. 12–14). Fig. 4 isolates the warp (4-step FID from 2.79 to 2.62 on CIFAR-10 and from 2.78 to 2.58 on ImageNet64), and Table 4 shows that replacing the learned clock with fixed schedules degrades FID at every step count.

Q1 (individual contributions).

We separate the two components with two controlled comparisons. First, Fig. 4 adds them one at a time on top of arbitrary-time distillation, with the number of updates and all other settings fixed:

4-step FIDLmatch\mathcal{L}_{\text{match}}+Ldef+ \mathcal{L}_{\text{def}}++ warp
CIFAR-102.952.792.62
ImageNet643.052.782.58

Turning on the warp, which sets both the training-time sampling and the inference grid, lowers the 4-step FID by a further 0.17 and 0.20. Second, Table 9 isolates inference-time scheduling: on one checkpoint, it changes only the inference grid. No fixed shift is best at every budget, and the learned grid gives the best 2-step FID and the lowest worst-case FID (3.24 vs. 3.29–3.56). Since one set of bin masses drives both components by design (Eqs. 12–14), Table 4 evaluates the warp as a whole and shows that replacing it with fixed schedules degrades FID at every step count.

Q2 (matched compute).

The extra cost is small: the defect adds student forward passes, and the warp update is an EMA over a few bins. The three runs in Fig. 4 share the same number of updates, so the gains of Ldef\mathcal{L}_{\text{def}} and the warp are measured at matched budgets. Against other distillation methods, DG-TWFD trains on a single A100 GPU, while CTM, for example, reports four V100 GPUs for CIFAR-10 and eight A100 GPUs for ImageNet64 and also trains a discriminator, so our budget is of the same order as the baselines we compare with.

Reviewer_8oNaOpenReview ↗

Thank you for your reply!

I thank the author for their detailed reply. I apologize for overlooking the ImageNet-256 experiments in Appendix F.2.. However, these results do not sufficiently address my concern for two reasons: first the FID is not evaluated at its standard protocol, and the result is not promising enough to demonstrate whether the method is practical.

I agree that not every suggested experiment is necessary, but I find that only experiment at small scale datasets like CIFAR10 and Imagenet64 does not provides much insight for demonstrating the effectiveness of the method.

Moreover, even if we are talking about small scale dataset experiments, many baselines in literature are neglected by the paper. Please see the references below.

Therefore, I am lowering my score.

[1] Xu, Yilun, Weili Nie, and Arash Vahdat. "One-step diffusion models with $ f $-divergence distribution matching." arXiv preprint arXiv:2502.15681 (2025).

[2] Geng, Zhengyang, et al. "Mean flows for one-step generative modeling." Advances in Neural Information Processing Systems 38 (2026): 75460-75482.

[3] Geng, Zhengyang, et al. "Consistency models made easy." International Conference on Learning Representations. Vol. 2025. 2025.

[4] Zhou, Mingyuan, et al. "Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation." Forty-first International Conference on Machine Learning. 2024.

[5] Zhou, Mingyuan, et al. "Adversarial score identity distillation: Rapidly surpassing the teacher in one step." arXiv preprint arXiv:2410.14919 (2024).

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

In this paper, the authors address the central challenge in few-step diffusion distillation: a student model trained on single-step transitions suffers from inconsistent outputs when its predictions are recursively composed across varying step budgets, a discrepancy termed the semigroup defect. To resolve this, the authors reformulate fast diffusion distillation as a compositional time-geometry adaptation problem and propose Defect-Guided Time-Warp Flow Distillation (DG-TWFD), which jointly learns a continuous-time student flow map together with a learnable monotone time reparameterization. Experiments on unconditional CIFAR-10 and class-conditional ImageNet64 demonstrate that DG-TWFD achieves improved FID scores, and supports arbitrary step counts from a single checkpoint without retraining.

Critical strengths

The paper addresses the important problem of compositional inconsistency in few-step diffusion distillation, where direct and recursively composed transitions may produce different results. By introducing a defect-guided monotone time warp, the method adaptively allocates training samples and inference steps to regions with larger compositional errors. This provides a reasonable approach to improving robustness across different inference step budgets while preserving the teacher's underlying trajectory.

Critical weaknesses
  • The proposed semigroup defect loss closely resembles existing consistency and trajectory-based distillation objectives, such as CTM [1], while adaptive timestep scheduling has been extensively studied [2, 3]. The main contribution appears to be combining established techniques using semigroup defects as an allocation criterion, without a sufficiently distinct methodological advance.

  • The experiments lack direct comparisons with existing adaptive timestep scheduling methods. Moreover, the main table reports CTM without post-training, whereas substantially stronger results have been reported for the full CTM framework [1]. These limitations make it difficult to establish whether the proposed method offers meaningful improvements over existing approaches.

  • The evaluation is limited to low-resolution toy benchmarks. Modern fast-diffusion research typically demonstrates scalability on high-resolution benchmarks such as 512x512 or 1024x1024.

References [1] Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion. ICLR, 2024. [2] Adaptive Time-Stepping Schedules for Diffusion Models. UAI, 2024. [3] Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training. CVPR, 2025.

Questions for authors

• What motivates using semigroup defects as the timestep allocation criterion? Since compositional consistency does not necessarily imply accurate approximation of teacher trajectories, is there empirical evidence that larger defects correlate with higher approximation errors or greater benefits from additional sampling steps? • Although the authors claim the objective is to jointly learn the time-wrapping and the distillation, it is unclear how they are being jointly learned. It misses a baseline of separating the time-warping and the distillation. Also, how about simply do the time-wrapping to ensure a uniform change in teacher states? • Could the authors clarify why CTM is evaluated without post-training and whether the reported performance comparisons use comparable training settings and computational budgets? • What are the additional training and storage costs associated with teacher trajectory caching, defect estimation, and time-warp optimization? Could the authors report the total training time, GPU memory usage, and storage requirements?

2 comments
AuthorsOpenReview ↗

Response to Reviewer TGbi (1/2): Weaknesses

We thank the reviewer for the thoughtful questions. We answer each point below.

W1 (novelty).

DG-TWFD introduces a closed loop in which the student's own composition error defines time. The semigroup defect, computed from the student alone, is minimized as a loss and, through one learned warp, decides where training triplets and inference steps are placed (Eqs. 12–14). Eq. (16) gives the reason: the multi-step error is bounded by the local errors plus exactly these defects, so the quantity we minimize is also the quantity that should allocate time. Prior work covers one piece each. CTM imposes the same agreement between direct and composed jumps with a fixed time distribution; [2] allocates the inference steps of a pre-trained model from its sampling error, and [3] allocates the training timesteps of a diffusion model trained from scratch from its loss decrease. None of them uses the composition error of a few-step student, and none couples training and inference through one schedule. This coupling yields the gains over both lines: over arbitrary-time distillation alone (Fig. 4) and over the tuned shifts and adaptive schedules compared in Tables 4 and 9. We will cite and discuss [2, 3].

W2 (comparison with adaptive schedules and with the full CTM).

[2] and [3] are not distillation methods. [2] chooses the sampling steps of a pre-trained multi-step diffusion model by minimizing a bound built from that model's score-matching error at each time; a few-step student that jumps directly between times has no such score error, so [2] does not apply to it as-is. [3] reweights timesteps when training a diffusion model from scratch, a different setting from distillation. We will discuss both in the related work. Within distillation, Table 4 compares the learned warp with two recent adaptive schedules (Pei et al., 2025; Stancevic et al., 2025) used as the student's time coordinate under the same framework, and the learned warp is best at every step count. As for the full CTM, it adds GAN post-training, which trains the student against real data with a discriminator. This stage goes beyond distilling the teacher, so its gains measure a different ingredient from the one our paper studies. DG-TWFD addresses the distillation itself: how a few-step student learns the teacher's transitions and where it spends its steps. The like-for-like comparison is therefore with CTM trained by its distillation objective alone, which is what the main table reports; the full CTM is listed in Table 12 for completeness.

W3 (resolution).

Appendix F.2 already scales DG-TWFD to 256×256256 \times 256 on latent ImageNet, where the warp lowers the defect by 6–12% at every checkpoint. This demonstrates scalability; 512×512512 \times 512 or 1024×10241024 \times 1024 use the same latent formulation and would not add further insight.

(Continued in the next comment: answers to the questions.)

AuthorsOpenReview ↗

Response to Reviewer TGbi (2/2): Questions

Overview.

These questions go to the idea behind DG-TWFD, which we state directly. A few-step generator succeeds or fails when it chains its own jumps, and the composition error it makes there, belongs to the student alone. Existing schedules place time by reading the teacher or by a fixed rule, so they cannot see this error. Our insight is that the student can measure this error by itself, without querying the teacher, and can therefore decide its own time: where it composes poorly, it needs more training and more steps. Each design choice follows from this one idea. The defect is a training loss because it is the error we want to remove; it allocates time because it locates where removal is hardest; and training and inference share one warp because the student composes the same jumps in both. Eq. (16) turns this idea into a bound, and Table 3 shows it in practice: two checkpoints with the same 1-step FID diverge as steps are added, and the one with the larger defect is nearly twice as bad at 4 steps.

Q1 (why the defect allocates time).

Fig. 2 shows the motivation: the defect is uneven along the trajectory, and the warp gives the high-defect regions more training samples and steps. By Eq. (16), the multi-step error is bounded by the local errors plus the defects, so the defect is part of the approximation error. Tables 3 and 12 confirm this: higher-defect models lose quality as steps are added (Table 3: 4-step FID 10.29 vs. 5.24; TCM: 2.08 at 2 steps to 2.87 at 8 steps), while low-defect models improve or stay stable.

Q2 (how the warp and the student are learned together).

The student and the warp are updated in the same loop: the student trains on triplets sampled under the current warp, and the warp is updated from EMA statistics of the defect and the teacher variation (Eqs. 10–13) together with a warp loss. Fig. 4 provides the separated baseline (Lmatch\mathcal{L}_{\text{match}} + Ldef\mathcal{L}_{\text{def}} without the warp). A warp that equalizes the teacher change is the special case that keeps only the teacher term VV of Eq. (12); it cannot see where the student fails, and Table 4 shows that such fixed schedules are worse than the learned clock.

Q3 (CTM without post-training and compute).

The full CTM adds a GAN loss on real data, which goes beyond distilling the teacher (see W2), so the main table compares DG-TWFD with CTM trained by its distillation objective alone. The training budgets are of the same order: each DG-TWFD run uses a single A100 GPU (Table 6), while CTM reports four V100 GPUs for CIFAR-10 and eight A100 GPUs for ImageNet64 (Kim et al., 2024, Appendix D), and DG-TWFD trains no discriminator. Within DG-TWFD, the components are compared at matched budgets: the three runs of Fig. 4 share the same number of updates.

Q4 (training and storage cost).

The CIFAR-10 teacher cache takes 13–16 GB of storage and is built once and shared by all runs (Table 7), so training needs no online teacher solves. The defect adds a few student forward passes per step, and the warp update is an EMA over a few bins with negligible cost. Every run, including the defect and warp terms, trains on a single 80 GB A100 GPU with fp16 mixed precision, which bounds its memory; total training time is reported in Table 6.

Reviewer_SJkhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies compositional consistency in few-step diffusion distillation: a student trained to predict individual transitions may behave differently when those transitions are recursively composed. The authors quantify this discrepancy through a semigroup defect and propose Defect-Guided Time-Warp Flow Distillation (DG-TWFD). The method combines cached teacher trajectories, a student transition map conditioned on source and target times, a normalized composition-consistency loss, and a monotone time warp driven by empirical defect and teacher-side variation. The same adaptive time allocation determines both training sampling and inference schedules.

The main experiments evaluate unconditional CIFAR-10 and class-conditional ImageNet64 using one checkpoint across 1, 2, 4, and 8 sampling steps. The reported eight-step FIDs are 1.81 and 2.49, respectively. Additional experiments examine loss components, prediction parameterizations, scheduling choices, cross-budget output deviations, sampling-seed variability, and preliminary transfer to DiT backbones.

Critical strengths
  1. A coherent connection between training allocation and inference scheduling. The method assigns more temporal mass to regions with larger empirical composition errors and uses this allocation both to sample training transitions and to place inference steps. This connects scheduling directly to the student’s modeling difficulties and constitutes the paper’s most compelling design choice.

  2. An explicit distinction between consistency and sample quality. Appendices C.4 and D examine cross-budget output deviations alongside FID and distance to a teacher reference solution. They also acknowledge that a smaller defect does not necessarily imply better FID at a fixed budget. This distinction helps clarify the intended benefit and provides a more informative assessment than reporting sample quality alone.

Critical weaknesses
  1. The experiments do not yet sufficiently isolate the contribution of defect-guided time warping. The anytime-versus-endpoint comparison in Table 3 changes the coverage of supervised target times, so it does not isolate the effect of the explicit semigroup-defect penalty. Figure 4 provides a complementary loss ablation, but its numerical outcomes and relationship to the final training recipe are insufficiently specified. Moreover, Appendix C.2 explicitly leaves open whether training-time warping outperforms a tuned inference-time shift, and defect-based grid selection does not identify the best two-step FID in Table 9. These results support the approach as a useful heuristic, but leave its incremental benefit over well-tuned scheduling and arbitrary-time transition distillation uncertain.

  2. The final performance results and the evidence for compositional robustness are not established using a consistent checkpoint and training protocol. Table 1 compares against CTM without post-training, whereas the official CTM evaluated in Table 12 achieves 1.91/1.86 FID at one/two steps and cross-budget deviations comparable to the evaluated DG-TWFD checkpoint. That DG-TWFD checkpoint is an earlier 61K checkpoint; the final checkpoint underlying Table 2 was not re-evaluated under this protocol. Appendix D also mentions an adversarial perceptual stage that is not fully specified in the main training objective. Finally, the main-run update counts in Table 6 are not clearly reconciled with the substantially larger budgets in Tables 10–11. Together, these issues limit attribution of the reported gains and reproducibility of the final results.

Questions for authors
  1. Could you provide a consolidated description of the training recipe for the final checkpoints in Table 2, including training stages, objectives, teacher checkpoints, total updates, and compute? In particular, how does the adversarial perceptual stage mentioned in Appendix D enter this recipe, and which runs do the budgets in Tables 6 and 10–11 describe?

  2. Could the final Table 2 checkpoint and official CTM be evaluated under a clearly specified, consistent protocol, reporting FID, cross-budget output deviation, and semigroup defect? Please also clarify how the learned continuous warp used for the main results relates to the integer-index Karras-grid protocol described in Appendix D.

  3. With arbitrary-time supervision, training budget, and all other training stages held fixed, what are the separate contributions of the explicit defect loss, adaptive training-time sampling, and adaptive inference scheduling? How does the full method compare with a tuned fixed timestep shift under these controls?

  4. Could you state the assumptions and provide a derivation for Equation (16), including error-propagation factors and why defects on adjacent triplets suffice to control the direct transition over a longer interval? For a fixed student, an invertible time reparameterization reindexes the defect at corresponding triplets. How does the argument distinguish this invariance from improvements caused by changes in training allocation, finite-capacity approximation, and discrete inference grids?

3 comments
AuthorsOpenReview ↗

Response to Reviewer SJkh (1/2): Weaknesses

We thank the reviewer for the detailed reading. We answer each point below.

W1 (isolating the contribution of defect-guided warping).

The benefit of DG-TWFD over both reference points is measured directly. Over arbitrary-time transition distillation alone (Lmatch\mathcal{L}_{\text{match}}), the defect loss and the warp lower the 4-step FID by 11% on CIFAR-10 and 15% on ImageNet64 (Fig. 4). Over well-tuned scheduling, the learned warp outperforms two recent adaptive schedules at every step count (Table 4) and has a lower worst-case FID than every tuned shift (Table 9). Table 3 tests the premise of our method: whether the semigroup defect tracks behavior across step counts. The two runs differ only in the target time; at 7K updates they reach the same 1-step FID (7.05 vs. 7.10), while the higher-defect run nearly doubles its 4-step FID (10.29 vs. 5.24). The effect of the defect loss itself is isolated in Fig. 4, the loss ablation with all other settings fixed: adding Ldef\mathcal{L}_{\text{def}} lowers the 4-step FID from 2.95 to 2.79 on CIFAR-10 and from 3.05 to 2.78 on ImageNet64, and the time warp lowers it further to 2.62 and 2.58. The ImageNet64 panel follows the final 400K-update recipe and ends at the Table 2 value; the CIFAR-10 panel uses the ablation setting of Table 4 for 300K updates and ends at the Table 4 value, and the caption will state this together with the endpoint values. Table 9 shows that no single fixed shift is best at every budget, and every shift is worse than our learned grid at 2 steps; the learned grid has the lowest worst-case FID (3.24 vs. 3.29 to 3.56), so one learned schedule serves all budgets without per-budget tuning. Finally, with two steps the grid has a single triplet, so the defect only measures how close the 2-step output is to the 1-step output, while FID measures quality (Appendix C.4); defect-based selection picks the FID-optimal grid at 4 and 8 steps, and we restrict the claim to these budgets.

W2 (consistent checkpoint and protocol).

DG-TWFD uses no adversarial post-training, so Table 1 compares it with CTM without post-training, the like-for-like setting under the official CTM protocol, and the full CTM with GAN is listed in Table 12 for reference. Following the reviewers' suggestion, we evaluated the final 400K checkpoint under the protocol of Table 12 and replace the 61K row, an early checkpoint of a GAN branch, with it (table below). Its cross-budget deviation is lower than that of IMM and TCM by factors of 2 to 7, CTM w/ GAN by around 1 and it is the flattest of all models: it grows by 24% from 2 to 8 steps, against 33% for CTM, 47% for IMM and fourfold for TCM. The adversarial stage is not part of DG-TWFD, and all main results use Eq. (9) only. It was tried on the 61K branch, where it raised the defect from 0.0240 to 0.0258 (Table 13), so we did not adopt it and report it only in the appendix; we will revise Appendix D to state this. Table 2 uses the 400K-update checkpoints, the same runs as the base row of Table 10. The update counts in Table 6 (50–70K and 100–200K for the main runs, 30–50K for the ablations) refer to intermediate checkpoints, used to compare compute with other methods at the 100-GPU-hour scale; the other rows of Tables 10–11 are exploratory scaling runs.

ModelFID (1 step)FID (2 steps)FID (4 steps)FID (8 steps)D(2)D(2)D(4)D(4)D(8)D(8)
CTM (official, with GAN)1.911.861.881.977.68.510.1
DG-TWFD (400K, final)2.62†2.05†1.84†1.81†7.07.58.7
IMM3.202.242.112.2045.157.566.1
TCM2.702.082.542.8721.767.288.4

Table 12 (updated): One checkpoint across budgets, under one protocol (CIFAR-10). FID: 50k samples, OpenAI evaluator. D(k)D(k): RMSE between the one-step and the kk-step output from the same noise, in gray levels (2048 samples). † Values from Table 2, evaluated on the learned warp grid of DG-TWFD.

(Continued in the next comment: answers to the questions.)

AuthorsOpenReview ↗

Response to Reviewer SJkh (2/2): Questions

Q1 (training recipe).

The Table 2 checkpoints are trained in a single stage without adversarial training, using the objective of Eq. (9) plus a warp loss that guides the time grid of Eqs. (10–13), the google/ddpm-cifar10-32 and EDM ImageNet64 teachers, and 400K updates on one A100 GPU; all hyperparameters are listed in Table 5. The adversarial stage in Appendix D is an exploratory branch outside this recipe. For Tables 6 and 10–11, see W2.

Q2 (evaluation protocol and grid).

All FIDs follow the official CTM protocol: 50k samples, the OpenAI evaluator and the CTM reference statistics. DG-TWFD is sampled on its learned warp grid (Eq. 14), while the integer-index Karras grid of Appendix D is the protocol prescribed for the baselines. The sentence in Appendix D that attaches this grid to Table 2 is a writing error and will be corrected. The updated Table 12 compares the 400K checkpoint with the official CTM under this protocol (FID and D(k)D(k)).

Q3 (separate contributions).

Fig. 4 keeps arbitrary-time supervision, the training budget and all other settings fixed and varies only the objective. The defect loss improves the 4-step FID by 0.16 (CIFAR-10) and 0.27 (ImageNet64), and the time warp by a further 0.17 and 0.20. One set of bin masses defines both the training sampling and the inference grid, so the warp is evaluated as one component. For tuned fixed shifts, see W1.

Q4 (assumptions and derivation of Eq. (16)).

Assume the student map is LL-Lipschitz in its state and the teacher flow satisfies the semigroup property. On a grid starting at u0u_0, let eke_k be the error of the direct jump from u0u_0 to uku_k, εk\varepsilon_k the local matching error on segment kk, and Δk\Delta_k the defect of the triplet (u0,uk,uk+1)(u_0, u_k, u_{k+1}). The triangle inequality gives

ek+1≤Δk+Lek+εk,e_{k+1} \le \Delta_k + L e_k + \varepsilon_k,

and unrolling the recursion gives

eK≤∑k=0K−1LK−1−k(εk+Δk).e_K \le \sum_{k=0}^{K-1} L^{K-1-k} (\varepsilon_k + \Delta_k).

This corrects Eq. (16): it adds the error-propagation factor and uses the triplets that share the start point, which our objective samples (Eq. 8). For a fixed student, a reparameterization leaves this bound unchanged; the warp acts through where training triplets are sampled and which grid points enter the sum.

Thank you for the reply. The training recipe (Q1), the consistent re-evaluation of the final 400K checkpoint (W2), and the component-level ablation in Fig. 4 are now answered. I maintain a rating of 2 (Weak Rejection), and four questions remain:

  • A tuned fixed-shift comparison. The paper claims the warp outperforms a tuned fixed shift, yet no table lists a tuned shift as a comparator under tuning effort equal to the proposed method (Fig. 4 compares against the same method without the warp; Table 4 against two other adaptive schedules). Could you report that comparison in one table, with FID, defect and cross-budget deviation together, and state the tuning budget used on both sides?
  • The margin in Table 9. The worst case is 3.24 versus 3.29, a difference of about 0.05 FID (1.5%), and the search cost behind each shift is not reported. Over how many seeds is this gap stable, and how does it compare with the search cost of your method?
  • The grids in Table 12. The † rows are evaluated on the learned warp grid while the baselines use the integer Karras grid; when the grids differ, "lowest and flattest deviation" is not a cross-method ordering. Could you re-run on a unified grid, or report both grids? Please also state the trade-off implied by 2.62 at one step and 2.05 at two against 1.91 and 1.86.
  • The three contributions and the scope of the claim. Q3 states that training sampling and the inference grid share one set of bin masses and cannot be attributed separately. Can the main text be narrowed accordingly, so that adaptive inference scheduling is no longer presented as an independently validated component? One decoupling run (training sampling on the warp, inference on the integer grid, and the reverse) would close this point.

Separately, the derivation of Eq. (16) is missing both formulas in the reply (nothing follows "the triangle inequality gives" and "unrolling the recursion gives"); please supply them so that the bound can be assessed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.