acceptodds
Under review as a conference paper at ICLR 2027

Variance Path Diffusion: Per-Region Noise Schedules for Inpainting and Motion In-Betweening

Abstract

Diffusion models corrupt each coordinate on a single, predefined noise schedule, yet most conditional tasks, such as inpainting, local editing, and motion in-betweening, require noise levels that differ across space and time. Existing methods design such region-wise schedules by hand, each rediscovering empirically the constraints a schedule must satisfy. We derive these constraints instead, by modeling the variance as a random variable: Variance Path Diffusion (VPD) models the cumulative noise as a Gamma process pinned in expectation to a nominal schedule, making variance allocation a free design variable. We prove that the natural shared-rate construction cannot meet a sufficient condition for well-behaved training at any finite stochasticity, and repair it with a Time-Varying-Rate (TVR) construction based on Beta thinning, which restores admissible VPD properties while leaving the marginals intact, and show that the randomness of the path is bounded and empirically does not affect quality. What remains is structure. A single allocation parameter , applied at sampling time to one model trained once, spans unconditional generation, editing, and inpainting. VPD covers temporal allocation for motion. VPD achieves the lowest or tied all-joint keyframe error on HumanML3D in-betweening, and sets a new state of the art for inpainting on FFHQ, LSUN Bedroom, and ImageNet.

Primary area:
generative models
Keywords:
diffusion, inpainting, motion in-betweening, per-region noise, generative model

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_gC3XOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Variance Path Diffusion (VPD) models cumulative diffusion variance using Gamma distributions and introduces a time-varying-rate construction based on Beta thinning. Regional variance allocation provides a single control parameter spanning generation, editing, inpainting, and motion in-betweening. Validation includes three image datasets, HumanML3D, and ablations of variance randomness and training mixtures.

Critical strengths
  • The Beta-thinning formulation preserves Gamma marginals and supports monotone variance paths with tractable conditional bridges. This provides a useful mathematical framework for designing regional noise schedules.

  • VPD improves most reported inpainting settings; for example, ImageNet wide-mask FID decreases from 56.7 to 38.5 relative to RAD. Motion experiments also show strong quality and keyframe accuracy, with baselines evaluated under matched keyframe spans. These results support practical value beyond the theoretical formulation.

Critical weaknesses
  • The shared-rate impossibility claim exceeds the demonstrated result. Corollary 3.2 bounds terminal variance variability when (A_1>1); it does not establish impossibility at every finite stochasticity parameter or concentration of the entire path. Moreover, (A_1>1) is a sufficient condition rather than a demonstrated requirement for successful training. This weakens a central motivation for the proposed construction.

  • VPD shows promising results on Table 2. However, the models requires a high NFE number and it is not clear if VPD works in one or few step generation (MeanFlow).

Questions for authors
  • What are the exact inside/outside variance schedules used for the reported results? Does Table 7 describe actual diffusion variance or a floor applied only to conditioning inputs? Please reconcile it with Appendix F.4.

  • Can the shared-rate claim be narrowed to the terminal-concentration tradeoff actually proved? What establishes that (A_1>1) is necessary for the practical training behavior being discussed?

  • Please provide the explanation if VPD can be extended to one or few step generation ?

2 comments

We thank Reviewer gC3X for their comments. The questions on the scope of the shared-rate claim and on the sampling schedules of Table 9 (prev. Table 7) and Appendix F.4 pointed to places where the paper was imprecise, and the question on few-step sampling led to a new set of experiments. We respond to each point below.

Q1: What is the exact inside/outside variance schedules? Clarify Table 7 and Appendix F.4

Table 9 (prev. Table 7) lists the inference settings used for all reported results. The floor σmin\sigma_{\mathrm{min}} is applied to the total per-pixel variance of the path itself, and the denoiser receives the same floored map. There is no separate floor on the conditioning input. Reconciling with App. F.4, each region's level is γrσˉt\gamma_r\bar\sigma_t with one fixed γ\gamma, clipped to [σmin,σmax][\sigma_{\mathrm{min}},\sigma_{\mathrm{max}}].

Q2: Can shared-rate claim be narrowed to the terminal-concentration tradeoff? Is A1>1A_1>1 necessary conditions?

We agree on both points. What we prove is a trade-off. Under a shared rate, A1>1A_1>1 forces the relative deviation of the terminal variance below σmin⁡/σmax⁡\sigma_{\min}/\sigma_{\max} (Prop. 3.1, Cor. 3.2). This does not show impossibility at every finite κ\kappa. Therefore, A1>1A_1>1 is also sufficient and not necessary. Intuitively, it ensures that every Gamma marginal has a density that vanishes at zero, so probability mass does not pile up at the floor. FID stays flat down after the condition (Fig. 2(c)), and what tracks the degradation is the fraction of levels clipped at σmin⁡\sigma_{\min} (Sec. 3.2). We have narrowed the claim in the abstract, contribution (2), and Section 3 (marked in green).

Q3: Can VPD be extended to one or few step generation?

VPD is a multi-step sampler and we have not tested one-step generation. One-step and few-step methods either distill a multi-step teacher (Progressive Distillation [1]) or train a few-step map directly (Consistency Models [2] or MeanFlow [3]). Our construction instead changes the noise-level law used in training and the per-region levels used in sampling. Therefore we think these methods can be applied with VPD formulation, with the per-pixel level map as an additional input. We leave this to future work.

Table. R4 Few-step sweep on FFHQ box masks (FID / LPIPS). DDNM and DPS use the same model as VPD TVR with a different sampler. AsyncPatch is our reimplementation, trained with its level distribution on the same model.

NFEVPD TVR (w/ Thm C.2)AsyncPatch (w/ Thm C.2)VPD TVR (w/o Thm C.2)AsyncPatch (w/o Thm C.2)DDNMDPS
425.98 / 0.1201153.83 / 0.313427.77 / 0.1042131.42 / 0.227847.56 / 0.165069.11 / 0.1998
823.11 / 0.073223.78 / 0.075225.26 / 0.080626.07 / 0.083427.72 / 0.112758.60 / 0.1669
1622.28 / 0.067323.11 / 0.071323.22 / 0.070824.35 / 0.075423.77 / 0.092650.71 / 0.1536
3221.86 / 0.065022.67 / 0.068922.24 / 0.066323.26 / 0.070923.24 / 0.081436.53 / 0.1238
6421.55 / 0.065122.25 / 0.068321.62 / 0.065422.52 / 0.069223.72 / 0.075726.78 / 0.0982

However, we provide the few-step sweep results in Table. R4. DDNM and DPS use the same denoiser as VPD TVR but differ only in the sampler. VPD's few-step behavior comes from conditioning through the level map together with training that covers it. The preserved region enters the network at σmin\sigma_{\mathrm{min}} from the first step, while DDNM and DPS see pure noise and impose the observation afterwards. AsyncPatch is our model trained with AsyncPatch's level distribution (as in Table 3), so it uses the same per-pixel level map and sampler. "w/ Thm C.2" uses the closed-form reverse covariance with the learned diagonal Σ0\Sigma_0 head and "w/o" uses the standard ancestral variance. AsyncPatch uses the same map and sampler but rarely trains on a region held at the floor, which breaks at NFE 4.

  1. [1] Tim Salimans and Jonathan Ho, "Progressive Distillation for Fast Sampling of Diffusion Models", 2022, ICLR
  2. [2] Yang Song et al., "Consistency Models", 2023, ICML,
  3. [3] Zhengyang Geng, et al., "Mean Flows for One-step Generative Modeling", 2025, NeurIPS

Dear Reviewer gC3X,

Thank you again for your careful review. We have posted our response and a revised manuscript. In particular, we narrowed the shared-rate claim to the trade-off that is proved, and added the few-step results.

As the discussion period closes soon, we would be grateful to hear whether the response addresses your concerns. We are happy to answer any further questions.

Best regards, Authors of Submission 48703

Reviewer_pJbwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces Variance Path Diffusion, a framework in which the variance of a diffusion process is taken as a random variable following a Gamma distribution. The authors study problems with the initial Gamma construction and propose a time-varying-rate formulation to address them.

They then apply this framework to per-region noise schedules, allowing the same approach to handle image inpainting, editing, and motion in-betweening. The method is validated experimentally on FFHQ, LSUN Bedroom, ImageNet, and HumanML3D, with better results in terms of performances and visually clear.

Critical strengths

The proposed method achieves good performance. The experiments provide good evidence of its practical value.

Interesting unified framework. Using per-region noise schedules to address different conditional generation problems is an interesting idea. I also found interesting the possibility of controlling the amount of editing with a single parameter.

Critical weaknesses

The motivation and significance of the theoretical construction are not fully clear to me. The Gamma distribution has some useful properties (stability, ...), but it is unclear why this distribution should be preferred over other ones. Additionally it seems to me that the experiments use deterministic variance paths, which makes me question how essential the stochastic formulation is to the practical method. Overall, I find the theoretical contributions interesting but somewhat limited in significance.

The presentation is too dense, and the relationship with existing methods could be clearer. In particular, in Section 5, it is not sufficiently clear whether eq (3) is a necessary constraint or a normalization chosen by the authors. I also think that the discussion of recent work on spatial scheduling and inpainting is incomplete. For example, Growing Images: Spatial Scheduling in Diffusion Inpainting (2026 FoGen workshop) appears relevant to this discussion.

Questions for authors

Is eq (3) constraint necessary for the construction to work, or is it a modeling choice to preserve the average variance?

Is there a reason to prefer Gamma distributions over other possible positive distributions?

4 comments

We thank Reviewer pJbw for their insightful comments. The fundamental question of why the Gamma law is preferred over other positive laws goes to the basis of our construction, and the remarks on presentation and related work have helped us make the paper clearer. The points you share with other reviewers (Eq. 3) are answered in the general response. We respond to each point below.

Q1: Is pin necessary? or just modeling choice to preserve the average variance?

It is a normalization, not a validity condition as addressed in the common concerns. Any pair of region levels can be written in coordinates (σˉ,γ)(\bar\sigma,\gamma) that satisfy Eq. 3, so Eq. 3 restricts no level map. Changing its right-hand side to a constant cc only rescales the schedule. What Eq. 3 fixes is the coordinate system in which we specify the training law. We also prove that the law does matter by replacing it by independent sampling of the two levels raises FID by 40–104% on motion with body-part masks. For more details, please check the common question section.

Q2: Why Gamma distributions to construct the stochastic pipeline?

A variance path needs positive, independent increments, and several laws provide these properties. We use the Gamma law because it is the only one with all three properties the construction relies on.

First, a sum of Gamma increments with a common rate is again Gamma, such that the marginal at every step is in closed form.

Second, the fraction of the total contributed by one increment is Beta distributed and independent of the total. By Lukacs' theorem, the Gamma law is the only one with this independence. It gives the reverse moments in closed form (Theorem C.2) and separates the total variance from its allocation over time (Appendix C.3).

Third, a Gamma variable multiplied by an independent Beta variable is again Gamma, which is what the time-varying rate of Section 4 uses.

Among the alternatives, stable laws, as used in LIM and DLPM, are closed under addition at any scale, but the corresponding variance has no finite mean, so it cannot be matched to a schedule in expectation. A Gamma law on the precision, as in t-EDM, gives a single draw shared by the whole trajectory and has no path structure in time (App. A).

Weakness 1: The role of the stochasticity

Table. R3 Stochastic (κ=5\kappa=5) versus near-deterministic (κ=5000\kappa=5000) training on FFHQ (FID / LPIPS). Δ\Delta is deterministic minus TVR.

NFEmaskVPD TVR (κ=5\kappa=5)deterministic (κ=5000\kappa=5000)Δ\DeltaFIDΔ\DeltaLPIPS
4box25.98 / 0.120127.26 / 0.1219+1.28+0.0018
4wide38.19 / 0.182540.61 / 0.1850+2.42+0.0025
4extreme96.28 / 0.4467102.77 / 0.4414+6.49−0.0053
8box23.11 / 0.073223.38 / 0.0737+0.27+0.0005
8wide24.14 / 0.117924.38 / 0.1190+0.24+0.0012
8extreme51.92 / 0.325452.47 / 0.3240+0.55−0.0013
16box22.28 / 0.067322.38 / 0.0677+0.11+0.0003
16wide22.77 / 0.109023.08 / 0.1104+0.31+0.0014
16extreme38.96 / 0.302340.30 / 0.3049+1.34+0.0026
32box21.86 / 0.065021.88 / 0.0654+0.02+0.0005
32wide21.99 / 0.105122.19 / 0.1059+0.20+0.0009
32extreme31.18 / 0.292632.09 / 0.2949+0.91+0.0023
64box21.55 / 0.065121.56 / 0.0649+0.00−0.0002
64wide21.98 / 0.105322.05 / 0.1060+0.06+0.0007
64extreme26.96 / 0.292427.53 / 0.2930+0.57+0.0005

On the role of the stochastic formulation, we first clarify the setup. Training uses the stochastic TVR path, and sampling uses its mean path. To measure how much the random path matters, we trained the same model on a deterministic path with the same budget and the same regional training distribution. At NFE 64 the two models differs by at most 0.6 FID. The stochastic model is better at few steps, by 1.3 to 6.5 FID at NFE 4 (Table R3).

We agree with the reviewer's reading. In the reported setting the practical method does not depend on the variance being random. The stochastic formulation contributes the analysis of Section 2 to 4, the separation of the total variance from its allocation that motivates Section 5, and the few-step gain above. We have revised the abstract, introduction, and section 8 accordingly.

Weakness 2: Missing recent literature

Thank you for the pointer. Growing Images paper studies the order in which masked regions are denoised at inference time with a fixed model, comparing parallel and progressive filling. This is complementary to our setting, where the model is trained on per-region noise levels and all pixels of the generated region follow one schedule. We have added Growing Images to Appendix A, together with other related works.

Dear Reviewer pJbw,

Thank you again for your careful review. We have posted our response and a revised manuscript. In particular, we answered why the Gamma law is used, clarified the role of the stochastic formulation with a new deterministic-path comparison, and revised the introduction and Section 8 accordingly.

As the discussion period closes soon, we would be grateful to hear whether the response addresses your concerns. We are happy to answer any further questions.

Best regards, Authors of Submission 48703

Reviewer_pJbwOpenReview ↗

Thank you for the response

Thank you for the detailed response, the revised manuscript and the new experiments. I find the comparison with deterministic-path training (Table R3) very useful.

My main concern remains. As stated in the rebuttal, in the reported setting the method does not depend on the variance being random: sampling uses the mean path, and on FFHQ a model trained on a near-deterministic path is within 0.6 FID at NFE 64 (the stochastic path helps only at very few steps). The per-region proposed construction of Section 5 is a design choice that does not need the Gamma law: Eq. (3) is a normalization, and the admissible ratio and start level (Prop. F.2, Cor. F.4) hold for a deterministic path as well. The Gamma-path analysis of Sections 2 to 4 is therefore separate from what produces the experimental results. The revised abstract and Section 8 now acknowledge this, but the title, the introduction, and the list of contributions still present the Gamma framework as the main contribution.

A related point is the comparison with RAD. In the inpainting experiments, the observed region stays at σmin⁡\sigma_{\min} for the whole chain, which is essentially the RAD setting. Table 3 shows that training only at this endpoint (pcor=1p_{\text{cor}} = 1) gives the same FFHQ FID as the full VPD (21.5 / 26.9 / 21.7 vs 21.6 / 26.7 / 21.7). Since the RAD numbers in Table 1 are copied from the RAD paper (correct me if I am wrong), the gain over RAD seems to come from implementation differences rather than from VPD. Retraining RAD in the same codebase, or stating this explicitly, would clarify this.

I think the paper would be stronger if it were framed around what the experiments support. A per-region parameterization with a single ratio γ\gamma would let one model move from generation to editing and inpainting.

Thank you for the follow-up and for the constructive suggestion.

Role of the Gamma framework. We agree on the facts you list, and our reading of them differs in one respect. Sections 2 to 4 are the analysis on which Section 5 is built. First, they answer whether a random noise level helps. Its effect is bounded (Prop. C.5) and appears only at very few steps (Table R3), which is what justifies sampling on the mean path. Second, they show that the path separates into a total and an allocation, and that scaling a region by a constant changes the rate and leaves the shape. The regional contrast can therefore be placed without affecting monotonicity, coarsening, or the reverse kernel (Prop. F.1). Placing it in the shape instead meets the collapse of Section 3 in every region (Appendix G). Section 5 follows this reading. It fixes the total by Eq. (3), which is a normalization as we state, and exposes the allocation as one ratio. The duality also says that the data constrain only the total, so the distribution of the allocation is a design variable, and our ablations measure its effect.

Title and contributions. The title names per-region noise schedules and the two tasks, which is the framing you suggest. In the contribution list, (1) and (2) are the analysis and (3) is the per-region allocation that the experiments evaluate. The revised abstract, introduction, and Section 8 state the role of each.

RAD. You are right that the RAD numbers in Table 1 are taken from the RAD paper. Our implementation is built on the RAD code base, so our inpainting endpoint is close to the RAD setting. Table 3 gives the comparison within our implementation, where endpoint-only training matches the full mixture on box and wide masks. In the motion comparison of Table 2, every baseline is re-evaluated by us under the same protocol.

What the experiments support. We agree, and thank you for pointing to the γ\gamma dial. The single ratio γ\gamma is what lets one model move between the three regimes, and the training distribution decides how well each is covered (contribution (3) in the manuscript). A model trained only at the endpoint, as in RAD, matches VPD on box and wide masks, but its FID is two to three times higher at intermediate ratios (Fig. 4). A model trained without endpoint mass covers the intermediate ratios well but is 3.6 to 6.9 FID worse on extreme masks (Table 3) and fails at NFE 4 (Table 11). Our mixture places mass on both and keeps the endpoint and the dial in one model.

Thank you once again for the valuable feedback.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Common Questions

We thank the reviewers for their careful reading. Two reviewers asked about Eq. (3), so we address it here. We then summarize the revisions to the paper.

About Eq. (3) (pin)

We agree that Eq. (3) (pin) is a normalization and section 5 in the paper now says so. We use two observations to make this precise.

First, replacing the right hand side of equation 3 by a constant cc changes the pixel-averaged variance to cσˉt2c\bar\sigma_t^2. This only rescales the nominal schedule by c\sqrt{c}, and the schedule's range is already a hyperparameter.

Second, drawing the two region levels σ1\sigma_1 and σ2\sigma_2 independently does not remove Eq. 3 either and just reparameterizes it. For any pair (σ1,σ2)(\sigma_1,\sigma_2), define σˉ2:=fσ12+(1−f)σ22\bar\sigma^2 := f\sigma_1^2+(1-f)\sigma_2^2 and γg:=σg/σˉ\gamma_g := \sigma_g/\bar\sigma. Eq. (3) then holds by construction.

Therefore, Eq. (3) does not restrict the level map but provides the coordinate system (σˉ,γ)(\bar\sigma,\gamma) in which we specify the training law where σˉ\bar\sigma follows the base model's law, and γ\gamma is 11, γcap\gamma_{\rm cap}, or log-uniform between them. The denoiser sees only per-pixel levels, so the real design choice is this joint density of the two levels which we ablate below.

We replaced the density by independent sampling of the two levels and kept everything else. B2 keeps the two atoms (point masses) at γ=1\gamma=1 and γ=γcap\gamma=\gamma_{\rm cap} and draws the two levels independently. B1 draws them independently with no atoms. For full-body keyframes (Table R1), keyframe error increases by 2 to 3% for B2 and by 7 to 11% for B1. FID moves in both directions across KK.

We then let the mask vary over the (frame, body part) grid, such that each keyframe specifies a random subset of body parts. The two regions and the sampling of γ\gamma are unchanged. Here the difference is large and consistent (Table R2). FID increases by 40 to 50% for B2 and by 82 to 104% for B1, and keyframe error by about 10%.

The training law we express through Eq. (3) is therefore a design choice with a measured effect, and the effect grows with the variety of masks.

Table R1. Training density ablation, motion in-betweening with full-body keyframes. Percentages are relative to pin.

training densityFID K=1K=1FID K=5K=5FID K=20K=20KF K=1K=1KF K=5K=5KF K=20K=20
pin (Eq. 3)0.10630.03800.03740.05710.03230.0167
B20.0913 (−14.1%)0.0388 (+2.1%)0.0354 (−5.3%)0.0585 (+2.5%)0.0333 (+3.1%)0.0170 (+1.8%)
B10.0947 (−10.9%)0.0557 (+46.6%)0.0361 (−3.5%)0.0632 (+10.7%)0.0358 (+10.8%)0.0178 (+6.6%)

Table R2. Training density ablation, motion in-betweening with random body-part keyframes. Percentages are relative to pin.

training densityFID K=5K=5FID K=20K=20FID randKF K=5K=5KF K=20K=20KF rand
pin (Eq. 3)0.06260.03320.04610.04690.02270.0382
B20.0939 (+50%)0.0466 (+40%)0.0682 (+48%)0.0519 (+11%)0.0249 (+10%)0.0421 (+10%)
B10.1221 (+95%)0.0604 (+82%)0.0942 (+104%)0.0526 (+12%)0.0249 (+10%)0.0427 (+12%)

Summary of revisions

All changes are marked in green in the revised PDF.

  1. We narrowed the claims on how the per-region construction relates to the stochastic framework, in the abstract, the introduction, Section 6 and 8, and the conclusion.
  2. Section 5.1 now states the Eq. (3) is a normalization, and Appendix F.1 gives the argument together with the new training-density ablations.
  3. The shared-rate result is stated as a trade-off, in the abstract, contribution (2), and Section 3.
  4. We added the comparison with deterministic-path training and the few-step results, and extended the related work in Appendix A.

We thank Reviewer Zn1P for their precise questions. The specific questions in regards to the theoretical constraints and empirical choices have made the paper clearer, and the questions about Eq. (3) led to a more careful discussion and to new ablations. We respond to each point below. The points you share with other reviewers are answered in the general response. We answer the rest here.

Q1: How should the theory be adjusted?

The implementation fixes one ratio γ\gamma for the whole chain and floors the cumulative variance of each pixel at σmin⁡2\sigma_{\min}^2. In the reported inpainting runs (γ=8000\gamma=8000 with σmax⁡=80\sigma_{\max}=80 and σmin⁡=0.01\sigma_{\min}=0.01), this places the observed region at the floor from the first step, so its level is in effect fixed at σmin⁡\sigma_{\min}.

The theory is adjusted as follows. The reverse bridge of Eq. (2) is well defined whenever each pixel's path is non-decreasing. Flooring is a monotone map, so the floored paths remain valid, and our sampler is the exact bridge of the floored path. It computes the removed fraction per pixel, w=1−Bt−1/Btw=1-B_{t-1}/B_t. While a region is at the floor, w=0w=0 and every reverse step leaves it unchanged. Only the final denoising call acts on it. The common removed fraction of Prop. F.1 holds on the steps where neither region is at the floor. It therefore describes the intermediate ratios of the editing sweep (Fig. 3(b)), and not the inpainting endpoint, where the observed region is at the floor throughout.

For training, we follow the stochastic TVR path with Eq. (3) and the admissible ratio of Proposition F.2. Since γ≤γcap(f,σˉt)\gamma\le\gamma_{\rm cap}(f,\bar\sigma_t), both mean region levels stay inside the trained range, so the adjustment concerns sampling only. The training hyperparameters are in Table 7 (previously Table 5).

Q2: How does deterministic-variance training perform?

Table. R3 Stochastic (κ=5\kappa=5) versus near-deterministic (κ=5000\kappa=5000) training on FFHQ (FID / LPIPS). Δ\Delta is deterministic minus TVR.

NFEmaskVPD TVR (κ=5\kappa=5)deterministic (κ=5000\kappa=5000)Δ\DeltaFIDΔ\DeltaLPIPS
4box25.98 / 0.120127.26 / 0.1219+1.28+0.0018
4wide38.19 / 0.182540.61 / 0.1850+2.42+0.0025
4extreme96.28 / 0.4467102.77 / 0.4414+6.49−0.0053
8box23.11 / 0.073223.38 / 0.0737+0.27+0.0005
8wide24.14 / 0.117924.38 / 0.1190+0.24+0.0012
8extreme51.92 / 0.325452.47 / 0.3240+0.55−0.0013
16box22.28 / 0.067322.38 / 0.0677+0.11+0.0003
16wide22.77 / 0.109023.08 / 0.1104+0.31+0.0014
16extreme38.96 / 0.302340.30 / 0.3049+1.34+0.0026
32box21.86 / 0.065021.88 / 0.0654+0.02+0.0005
32wide21.99 / 0.105122.19 / 0.1059+0.20+0.0009
32extreme31.18 / 0.292632.09 / 0.2949+0.91+0.0023
64box21.55 / 0.065121.56 / 0.0649+0.00−0.0002
64wide21.98 / 0.105322.05 / 0.1060+0.06+0.0007
64extreme26.96 / 0.292427.53 / 0.2930+0.57+0.0005

We agree that the main results rest on the regional allocation and the training mixture. To isolate path randomness in training, we retrained the FFHQ model with a near-deterministic path (κ=5000\kappa = 5000, RSD[BT]=κ−1/2≈1.4%{\rm RSD}[B_T]=\kappa^{-1/2}\approx 1.4\%), as shown in Table R3. The regional allocation, γ\gamma mixture, mask law, training steps and sampler are identical to VPD (κ=5\kappa=5). Both models are sampled on the mean path with the same sampler. At the reported setting in the main paper, the two models are indistinguishable.

The stochastic path helps at NFE 4. This is consistent with the rest of the paper. Proposition C.5 bounds the effect of the path randomness in sampling, and Table 6 (prev. Table 4) confirms it. The denoiser is conditioned on the realized level, so randomness in training acts only through the density of levels the denoiser sees where a random path broadens that density around the mean path. The new comparison shows the same for training at the reported setting. The stochastic formulation contributes the analysis of Sections 3 and 4, and the few-step gain above.

Q3: What are the new provable results compared to AsyncPatch?

AsyncPatch proves that jointly trained per-token diffusions recover the joint score (Lemma 1), and extends this to an arbitrary joint law over per-token times (Lemma 2), and gives an ELBO obtained by averaging over monotone paths (Theorem 1). In all three, the noise level of a token is a deterministic function of its time. Our mean-path sampler is of this type. Its validity follows from their Lemma 2, and we claim no new result there. For a fixed deterministic path, the bound is the classical DDPM decomposition along that path, which is also the building block of their proof.

Our new provable results concern a different object, a variance path that is itself random, plus closed-form admissibility for the per-region construction. Under a shared rate, the endpoint ratio of the shape parameter is fixed by the schedule, which gives a trade-off between randomness and clipping (Prop. 3.1, Cor. 3.2). Under a time-varying rate we prove exact Gamma marginals, the monotonicity condition, and exact coarsening on any subgrid (Props. 4.2-4.4). We give the reverse moments under a random path with closed-form coefficients, a bound on the total size of the path-uncertainty term (Thm. C.2, Prop. C.5), and an ELBO with the path as a latent variable (Thm. C.4). For the per-region setting, monotonicity, coarsening, and the reverse moments hold in each region when the path is rescaled by a constant (Prop. F.1). We also give the admissible range of the ratio γ\gamma and the start level in closed form (Prop. F.2, Cor. F.4), and these hold for a deterministic path as well.

Q4: Which are general validity conditions, and which are properties specific to the proposed path family or empirically motivated design choices?

  1. General validity for any variance path. The path must be non-decreasing. Given the path, the forward marginal is Gaussian and the reverse transition is a Brownian bridge (Eq. 2).

  2. Specific to the Gamma path. The endpoint bound under a shared rate (Prop. 3.1, Cor. 3.2). Exact marginals, the monotonicity condition, and exact coarsening under a time-varying rate (Prop. 4. to 4.4). The reverse moments and the bound on the path term (Thm. C.2, Prop. C.5).

  3. Design choices and their closed form consequences. Eq. (3) is a normalization, and γ\gamma is the coordinate we use for the two region levels (detailed at common comments above). Given these, the largest admissible ratio γcap\gamma_{\mathrm{cap}}, the saturation of the generated region, and the start level follow in closed form (Prop. F.2, F.3, Cor. F.4).

  4. Empirical choices. The training mixture of γ=1,γ=γcap\gamma=1,\gamma=\gamma_{\mathrm{cap}} and the log-uniform law between them. They are supported by Table 3, Fig. 2(c), Table 6 (prev. Table 4) and the new ablation above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.