acceptodds
Under review as a conference paper at ICLR 2027

Budgeted Block–Time Cache Scheduling for Diffusion Transformers

Abstract

Diffusion models have demonstrated remarkable performance in image and video generation, while their multi-step denoising process incurs substantial computational overhead. Feature caching accelerates sampling by reusing intermediate activations across timesteps, making the caching schedule critical to both efficiency and generation quality. Existing scheduling methods primarily operate at the timestep level, imposing a shared execution decision on all Transformer blocks within a timestep and thereby overlooking the heterogeneous cache sensitivity and computational redundancy across network depth. We propose BTCache, which models cacheability jointly across denoising steps and Transformer blocks. By predicting the quality impact of block–time reuse, BTCache identifies critical computations and allocates a prescribed compute budget across the joint block–time space. At each denoising step, a deterministic selector executes the budget-feasible action of lowest predicted deviation from full computation, selectively recomputing sensitive blocks while reusing redundant features within the same timestep. Extensive experiments on FLUX, HunyuanVideo, and Wan2.1 demonstrate that BTCache preserves high visual quality under prescribed compute budgets.

Primary area:
generative models
Keywords:
Diffusion Transformers.Inference Acceleration.

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_q8jpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BTCache allocates a fixed computation budget jointly across transformer block groups and denoising steps. It combines residual reuse with damped extrapolation, a learned ensemble that predicts terminal latent, perceptual, and temporal risks, and an exact feasibility test that preserves the remaining budget and maximum cache age. Training uses complete continuation trajectories and policy-visited branches rather than only local feature errors. Experiments on HunyuanVideo, Wan2.1, and FLUX report roughly 3.4–3.6× end-to-end speedups. The idea is useful, but the evidence does not yet clearly isolate the benefit of joint block decisions or validate the terminal-risk predictor that drives them.

Critical strengths
  1. Explicit control of the computation budget. Proposition 1 and the feasible-action filter give a concrete way to enforce both the exact number of remaining FULL cells and cache-age limits. This is more useful for predictable resource allocation than a threshold that only controls average compute, and the paper appropriately reports wall-clock measurements separately.
  2. Supervision targets the eventual generation outcome. The same-prefix branching scheme and separate terminal-risk heads address error accumulation and tradeoffs that a local feature-distance proxy can miss. In Table 4, replacing terminal supervision with local-error supervision reduces HunyuanVideo PSNR from 17.23 to 16.54 dB, supporting this design choice.
  3. The implementation includes realistic overhead measurements. The main latencies include the controller, and Appendix D reports controller costs of 1.74 s on HunyuanVideo and 1.08 s on Wan2.1. Tests span image and video generation, and the matched-cost FLUX comparison improves PSNR over MagCache while preserving similar LPIPS. These results show that controller overhead does not automatically erase the savings.
Critical weaknesses
  1. The evaluation does not yet establish a robust quality advantage. Section 4.1 uses only nine HunyuanVideo prompts and eight Wan2.1 prompt–seed pairs. Several quality margins are small: HunyuanVideo VBench is 79.44 versus 79.43 for TaylorSeer, without paired uncertainty estimates. The seven-of-nine win count and budget sweep support the internal comparison but do not establish broader generalization. On 40 FLUX prompts, Table 3 reports 1.47 dB higher PSNR than matched-cost MagCache with identical LPIPS (0.3218) and nearly identical SSIM. The faster MagCache setting also has higher PSNR but worse LPIPS than its published setting. These patterns are not intrinsically inconsistent, but the missing FLUX operating parameters and per-sample results prevent checking their cause. Broader held-out evaluation and verification of this comparison are needed before drawing a strong quality-lead conclusion.
  2. The closest scheduling comparisons are missing. Section 2 explicitly identifies MeanCache and DPCache as the closest formulations, yet neither appears in the experiments. Table 4 and Appendices D–E do support finer allocation under the authors’ controller; they do not establish superiority over these independent fixed-budget schedulers. Figure 1d also provides evidence that one-step reuse errors are nonadditive, but it does not measure the benefit of learning joint interactions over an additive scorer with the same feasible partial-block masks. A budget or latency sweep against an applicable close scheduler, together with this controlled additive comparison, would test the claimed scheduling advance more directly. Any incompatibility should be explained and the comparative claim scoped accordingly.
  3. The data requirements and offline cost of the learned controller cannot be assessed. Section 3.5 describes the rollout procedure and Appendix C specifies the architecture, so training is not wholly undocumented. However, the paper does not report training prompt sources and counts, terminal rollout and branch counts, or total data-generation, fitting, and calibration cost per backbone. It specifies calibration/test separation without fully describing the training split’s relationship to those sets. These omissions matter for a five-member ensemble trained on terminal outcomes: they prevent assessing data dependence, generalization, and the number of generations needed to amortize preparation. The reported 2% online controller overhead does not resolve this cost question.
Questions for authors
  1. How does the joint-mask predictor compare with an additive per-group predictor using the same feasible masks, terminal supervision, residual operator, and budget? This comparison would separate the value of modeling group interactions from the benefit of allowing partial refreshes.
  2. On held-out policy-visited states, can you compare predicted action rankings with terminal outcomes obtained by branching and then following the deployed continuation? Please report the selected-action regret and explain whether another policy-evaluation or data-collection iteration changes the decisions.
  3. How does BTCache compare with a closely related fine-grained scheduler at the same computation budget and approximation policy? Please include the offline training and calibration cost when assessing whether the additional learned machinery is worthwhile.
2 comments

We thank the reviewer for the thorough review and for recognizing the explicit budget control and the supervision on terminal outcomes. Because the questions were specific, we could run most of them as stated. These include the additive scorer under identical masks, ranking and regret on held-out policy-visited states, and a budget sweep against close schedulers with offline cost. Some of the results lead us to state our claims more narrowly, and each answer says where.

Response to Weakness 1

We agree that margins of this size need paired uncertainty estimates. As an additional evaluation, we extended HunyuanVideo to 36 test prompts at 12 NFE (48 of the 200 group–step cells; one timestep is 4 cells), one fixed seed per prompt. The prompts are the paper's 9 test prompts, 15 fixed VBench prompts and 12 further VBench prompts. We compare with MagCache configured to compute exactly 12 steps (MagCache-12; the stronger of two such settings, schedule {0–6, 13, 19, 28, 37, 46}). We evaluate BTCache; the reference plan π0\pi_0 re-selected after a second round of data on the training prompts

HunyuanVideoPSNR↑SSIM↑LPIPS↓
MagCache-1218.460.64560.401
TeaCache 0.30 (mean 12.31 NFE)17.100.62520.413
BWCache18.490.64570.3901
BTCache18.570.65190.3816

As an additional evaluation, we ran MagCache with its official forward pass on the 40 DrawBench prompts, with one fixed seed per prompt and the same all-full references as BTCache. In these runs BTCache has higher PSNR and SSIM and lower LPIPS than MagCache at every setting below.

BudgetMethodPSNR↑SSIM↑LPIPS↓
8 NFEBTCache24.240.73620.3218
MagCache22.770.73640.3218
TeaCache18.050.73560.3221
TaylorSeer16.090.64470.4498

Response to Weakness 2 & Question 1

Thank you for suggesting this controlled comparison. We trained an additive per-group predictor, which is the joint predictor with the cross-group interaction removed, on identical training episodes and with the same recipe, reference plan, β\beta, selector, budget and residual operator. The additive predictor gives lower PSNR and SSIM and higher LPIPS than the joint predictor.

HunyuanVideoPSNR↑SSIM↑LPIPS↓
Joint predictor (BTCache)17.230.6130.410
Additive predictor16.800.58380.4775

Response to Weakness 2 & Question 3

we ran a budget sweep against DPCache and BudCache on FLUX.1-dev, compared with BWCache on HunyuanVideo, and now state our comparative claims more narrowly. BTCache has the best PSNR, SSIM and LPIPS.

NFEMethodPSNR↑SSIM↑LPIPS↓
8BTCache24.240.73620.3218
8BudCache24.170.73110.3257
8DPCache19.260.76350.2919
11BTCache24.380.87310.1427
11BudCache23.070.83460.2012
11DPCache21.740.83030.1907
14BTCache27.880.90860.0986
14BudCache25.690.88230.1333
14DPCache23.630.87310.1318
20BTCache30.370.94180.0528
20BudCache28.510.92600.0727
20DPCache26.360.90860.0887

BudCache's official search takes 26 min (8 NFE) and 43 min (14 NFE) on our hardware. ECAD's released schedules sample 20 steps against our 50-step reference. BWCache's official code covers Open-Sora, with Wan and HunyuanVideo listed as future work, so we re-implemented it. MeanCache's official code covers Qwen-Image and Z-Image, and ProCache's covers DiT, FLUX and PixArt; we could not run either in the time available. DPCache's official code covers FLUX and Wan2.1-I2V-14B-720P only, and porting it to HunyuanVideo and Wan2.1-T2V-1.3B with its PACT calibration could not be completed in the rebuttal period.

Response to Weakness 3

We agree that the data requirements and offline cost belong in the paper, and App. I now reports them.

HunyuanVideo itemValue
RolloutsRound 1: 1200(30 prompts × 10 families × 4), about 180 GPU-h. Round 2: 60, about 9 GPU-h; about 27 GPU-h together. A800 with CPU offload; about 8 min per rollout, 23 min per all-full reference
Predictor5 members, independent seeds; AdamW, lr 8e-4, weight decay 1e-4, clip 5.0; one family per batch with full recurrent backprop; 600 updates per member in round 1, 900 in round 2; CPU, about 35 min in round 1
LossGaussian NLL on standardized log terminal risk ×1.0, same-state pairwise rank ×0.5, local NLL ×0.25, next-state smooth-L1 ×0.25
Comparator5 members, same optimizer

We fixed 12 states visited by the selector on HunyuanVideo before seeing any result. At each state we forced each of the 16 masks , followed BTCache to the final step, and scored PSNR against the all-full video. Repeating the selector's own branch reproduces it exactly in 10 of 12 states. Across the 16 masks, PSNR spans 2.14 dB on average.

State123456789101112
Best mask − best timestep-level mask (dB)0.2300.090.20000.190.1800.400.120.05
Partial masks better than best timestep-level (of 14)1031000860132
Regret of the executed mask (dB)000.090000.190.1800.0.080.05

To measure how much a better predictor could gain, we took 25 states (the 12 above and 13 round-3 states) and compared the best of the 16 masks with the best timestep-level mask under the same exact-budget continuation.

Metric (states)MeanMedianPartial mask best in
PSNR (25)0.20 dB (max 1.97)0.085 dB82%
LPIPS (25)0.10460.000480%
Temporal (13)0.1930.06175%
Reviewer_VhmoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes BTCache, which allocates a fixed compute budget across denoising steps and Transformer block groups. A learned predictor scores joint recompute/cache actions by terminal quality risk, while a feasibility rule enforces the exact budget.

Critical strengths
  • Predicting terminal quality loss is well motivated, and the ablation shows an advantage over local-error supervision.

  • The paper includes equal-budget ablations and counts controller overhead in the reported latency.

Critical weaknesses
  • The benefit of the main contribution is small. In Table 4, joint block–time allocation improves PSNR by only 0.15 dB and LPIPS by 0.005 over timestep-only allocation, at essentially the same speed. The ablations show a larger effect from terminal supervision than from depth allocation. A reference-plan-only baseline is also missing, leaving the value of online selection unclear.

  • The closest related methods, including MeanCache, DPCache, BWCache, and ProCache, are discussed but not compared experimentally. Without these comparisons under matched budgets, the paper does not establish its advance over existing budgeted or block-level scheduling.

  • Video evaluation uses only nine prompts for HunyuanVideo and eight prompt–seed pairs for Wan2.1, without uncertainty estimates. This is weak evidence for small gains such as the 0.0009 LPIPS improvement over TeaCache on Wan2.1.

  • Appendix C specifies the predictor architecture, but training prompt and trajectory counts, optimization settings, and terminal-rollout generation costs are missing. Since training is required for each backbone, these omissions prevent assessing both reproducibility and the practical cost of the method.

Questions for authors

See weakness for details.

1 comment

We thank the reviewer for the detailed and constructive review, and for the positive comments on the terminal-quality objective and the equal-budget ablations. Following the reviewer's suggestion, we added the reference-plan-only baseline; its result changes how we present the online selector.

Response to Weakness 1

We agree that the joint-allocation gain in Table 4 is small and that a reference-plan-only baseline was needed. As an additional evaluation, we added the reference-plan-only baseline on HunyuanVideo at 12 NFE (48 of the 200 group–step cells; one timestep is 4 cells) on 36 test prompts, one fixed seed per prompt. Bold marks the best value per column.

The reference plan π0\pi_0 (Sec. 3.6) is the fixed timestep-level plan toward which β\beta pulls the selector. BTCache improves on π0\pi_0.

Method (HunyuanVideo, 12 NFE)PromptsPSNR↑SSIM↑LPIPS↓
Reference plan π0\pi_03617.190.60050.4575
BTCache3617.230.6130.410

Response to Weakness 2

Thank you for this suggestion. Within the rebuttal period we compared our method with DPCache and BudCache on FLUX.1-dev and with BWCache on HunyuanVideo, all at matched budget; MeanCache and ProCache could not be run. BudCache, which the paper cites in Sec. 2, searches a fixed-NFE refresh mask offline. This additional evaluation uses FLUX.1-dev on the 40 DrawBench prompts, one fixed seed per prompt shared by all methods, at the paper's 8 NFE and at 11, 14 and 20 NFE. BudCache uses its official search and DPCache its official code with its released cost tensors. Bold marks the best value per column at each budget. BTCache has the best PSNR, SSIM and LPIPS.

NFEMethodPSNR↑SSIM↑LPIPS↓
8BTCache24.240.73620.3218
8BudCache24.170.73110.3257
8DPCache19.260.76350.2919
11BTCache24.380.87310.1427
11BudCache23.070.83460.2012
11DPCache21.740.83030.1907
14BTCache27.880.90860.0986
14BudCache25.690.88230.1333
14DPCache23.630.87310.1318
20BTCache30.370.94180.0528
20BudCache28.510.92600.0727
20DPCache26.360.90860.0887

BWCache's official code covers Open-Sora, with Wan and HunyuanVideo listed as future work, so we re-implemented it.

MeanCache's official code covers Qwen-Image and Z-Image, so it could not be run on our backbones. ProCache's official code covers DiT, FLUX and PixArt, and we could not complete a FLUX run in the time available. DPCache's official code covers FLUX and Wan2.1-I2V-14B-720P only, and porting it to HunyuanVideo and Wan2.1-T2V-1.3B with its PACT calibration could not be completed in the rebuttal period.

Response to Weakness 3

This is a fair concern, and we enlarged both video test sets. On 24 Wan2.1 prompt–seed pairs at 12 NFE, BTCache has higher PSNR and lower LPIPS than TeaCache in less time, and higher PSNR and lower LPIPS than the best 12-step MagCache at matched time. The pairs are the 12 prompts of our own evaluation list, each with 2 seeds, disjoint from the calibration prompts. MagCache-12 is MagCache configured to compute exactly 12 steps (best of four settings).

Wan2.1, 24 pairs, 12 NFEPSNR↑SSIM↑LPIPS↓
BTCache18.590.68330.2975
MagCache-1218.450.67730.3013
TeaCache 0.2518.030.67990.3011

Response to Weakness 4

We agree that these numbers are needed to judge reproducibility and cost. Only HunyuanVideo has a learned selector: 1260 rollouts on 30 training prompts, about 180 A800 GPU-h of collection, and CPU fitting. Each rollout is continued to the final step so that it carries terminal labels. The predictor is a 5-member ensemble trained with AdamW (lr 8e-4, weight decay 1e-4, clip 5.0), 600 updates per member in round 1 and 900 in round 2, about 35 min on CPU in round 1; the loss combines Gaussian NLL on terminal risk (×1.0), same-state ranking (×0.5), local NLL (×0.25) and next-state smooth-L1 (×0.25).

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

BTCache allocates a fixed computation budget across both denoising steps and Transformer groups. A learned predictor estimates the terminal latent, perceptual, and temporal errors of joint caching actions using information available along the accelerated trajectory. A deterministic selector then chooses an action subject to exact remaining-budget feasibility and a preference for a calibrated reference schedule. Experiments on HunyuanVideo, Wan2.1, and FLUX.1-dev report end-to-end speedups of 3.62×, 3.58×, and 3.37×, respectively.

Critical strengths

A Clear Formulation with Exact Budget Control.** Allocating computation across both network depth and denoising time is a useful approach, and the method considers joint caching decisions rather than treating each group independently. The feasibility condition has a clear proof and gives the scheduler a practical way to satisfy both the computation budget and cache-age constraints.

Promising Speedups with Useful Component Analysis.** The paper evaluates three backbones and reports full generation latency, including controller overhead. The controller accounts for about 2% of runtime on the video models. Controlled ablations examine action granularity, terminal supervision, state correction, and residual extrapolation, providing useful evidence about which parts of the method contribute to the results.

Critical weaknesses

Limited Empirical Evidence for the Scheduling Contribution.** The paper identifies MeanCache and DPCache as its closest scheduling formulations, but neither appears in the experiments. It also discusses ECAD as a fine-grained scheduling approach without comparing against it. DPCache and ECAD would be particularly useful comparisons because they test whether BTCache improves on strong temporal scheduling and offline fine-grained scheduling, respectively. The internal timestep-level ablation is helpful, but the gain from finer allocation is modest: on HunyuanVideo, VBench increases from 79.40 to 79.44, PSNR from 17.08 to 17.23, and LPIPS improves from 0.415 to 0.410. These comparisons use only nine test prompts. The other main results use eight prompt-seed pairs on Wan2.1 and 40 prompts on FLUX. With this limited coverage and no comparison to the closest scheduling methods, I am not yet convinced that the additional scheduling machinery offers a consistent advantage.

Insufficiently Specified Predictor Training and Bootstrap Procedures.** Section 3.5 and Appendix C describe the architecture and supervision, but omit the source and size of the predictor training set, the number of sampled states and terminal branches, and the optimizer, learning rate, and training duration. Section 4.1 lists calibration and test-set sizes, while the predictor's training split remains unclear. The bootstrap procedure also needs more detail: the predicted action values depend on the continuation policy used to produce terminal labels, so the sequence of rollout collection, predictor fitting, and selector updates affects what the predictor learns. These details, together with the total data-generation and training cost, are needed to reproduce the method and assess the effort required to adapt it to another backbone.

Questions for authors

Q1. Could the authors compare BTCache with DPCache and a fine-grained scheduling method such as ECAD on shared backbones at comparable measured latency? A few operating points would help establish where BTCache offers a better tradeoff.

Q2. What are the complete predictor training and rollout-collection procedures, including training data and splits, sampled state and branch counts, continuation policies, optimization settings, and total GPU-hours?

Q3. Do the gains hold on a larger, explicitly documented held-out prompt set? In particular, does the block-time versus timestep-level improvement also hold on Wan2.1 and FLUX? These results would help determine whether the benefit extends beyond the small HunyuanVideo ablation set.

3 comments

We sincerely thank the reviewer for the careful review and for recognizing the exact budget control with its feasibility analysis. We have run additional experiments and revised the manuscript accordingly; our responses follow the order of the review.

Response to Weakness 1 & Question 1 (closest schedulers)

We now compare our method with DPCache and BudCache on FLUX.1-dev at measured speedup, at the paper's budget and at three larger ones. At 8 NFE BudCache has the highest PSNR and speed; at 11, 14 and 20 NFE BTCache has the best PSNR, SSIM and LPIPS. ECAD and MeanCache could not be run in a comparable setting.

This evaluation was added for the rebuttal. It uses FLUX.1-dev with the 40 DrawBench prompts, one fixed seed per prompt shared by all methods, and metrics against the same-seed all-full image. The budgets are 8 NFE, the paper's setting (32 of the 200 group–step cells; one timestep is 4 cells), and 11, 14 and 20 NFE for the operating points requested in Question 1. BudCache, which the paper cites in Sec. 2, searches a fixed-NFE refresh mask offline; we use its official search. DPCache runs its official code with its released cost tensors. Bold marks the best value per column at each budget.

NFEMethodPSNR↑SSIM↑LPIPS↓
8BTCache24.240.73620.3218
8BudCache24.170.73110.3257
8DPCache19.260.76350.2919
11BTCache24.380.87310.1427
11BudCache23.070.83460.2012
11DPCache21.740.83030.1907
14BTCache27.880.90860.0986
14BudCache25.690.88230.1333
14DPCache23.630.87310.1318
20BTCache30.370.94180.0528
20BudCache28.510.92600.0727
20DPCache26.360.90860.0887

We could not run ECAD in a comparable setting within the rebuttal period. Its released FLUX schedules sample 20 steps, whereas our reference uses 50, and we did not run its search in our setting. MeanCache's official code covers Qwen-Image and Z-Image. DPCache's official code covers FLUX and Wan2.1-I2V-14B-720P only, and porting it to HunyuanVideo and Wan2.1-T2V-1.3B with its PACT calibration could not be completed in the rebuttal period.

Response to Weakness 1 & Question 3

We agree with the reviewer that the gain from finer allocation is modest. To answer Question 3, we compared block–time and timestep-level plans at the same budget on held-out sets for all three backbones: 24 HunyuanVideo prompts, 24 Wan2.1 prompt–seed pairs and the 80 DrawBench prompts on FLUX.1-dev.

The video test sets are larger and fully documented; On HunyuanVideo, the BTCache selector evaluated here was retrained during the rebuttal on the calibration prompts of Sec. 4.1. The 36 test prompts are disjoint from them: the paper's 9 test prompts, 15 prompts taken from fixed rows of the VBench prompt list before any result was seen, and 12 further VBench prompts disjoint from all others. The 24 Wan2.1 pairs are the 12 prompts of our own evaluation list, each with 2 seeds, disjoint from the Wan2.1 calibration prompts. Every prompt or pair has one fixed seed shared by all methods.

BackbonePlanPSNR↑SSIM↑LPIPS↓
HunyuanVideoBlock–time plan17.830.68110.381
Timestep-level plan17.310.65130.3932
FLUXTimestep-level plan19.200.78140.3365
Block–time plan24.280.78730.3205
Wan2.1Timestep-level plan18.330.67970.3002
Block–time plan20.280.72730.3205

Response to Weakness 2 & Question 2

We thank the reviewer for pointing out that these details were missing from the submission. We give them in full below. For the submission, we generated rollouts on 30 training prompts disjoint from the paper's 9 test prompts and continued each to the final step, so that it carries terminal quality labels. Rollouts come in families: a base schedule plus three branches that leave it at later steps; each prompt has 10 families of 4 rollouts, 1200 rollouts in total, collected in a single round. We then fit a 5-member predictor ensemble and a comparator network that predicts the terminal risk of the reference plan. Finally we choose the reference plan π0\pi_0 and the weight β\beta on the calibration prompts of Sec. 4.1; π0\pi_0 is the timestep-level base schedule with the lowest mean predicted risk, each schedule scored by the largest of its latent, LPIPS and temporal risks.

HunyuanVideo itemValue
RolloutsRound 1: 1200(30 prompts × 10 families × 4), about 180 GPU-h. Round 2: 60, about 9 GPU-h; about 27 GPU-h together. A800 with CPU offload; about 8 min per rollout, 23 min per all-full reference
Predictor5 members, independent seeds; AdamW, lr 8e-4, weight decay 1e-4, clip 5.0; one family per batch with full recurrent backprop; 600 updates per member in round 1, 900 in round 2; CPU, about 35 min in round 1
LossGaussian NLL on standardized log terminal risk ×1.0, same-state pairwise rank ×0.5, local NLL ×0.25, next-state smooth-L1 ×0.25
Comparator5 members, same optimizer

Dear Authors,

I appreciate the clear budget formulation and the reported end-to-end speedups. My main concern remains whether joint block and timestep scheduling offers a consistent quality advantage over closely related schedulers at comparable measured latency. The internal ablation shows a modest gain. A comparison with DPCache or ECAD on a shared backbone would help establish the practical value of the proposed scheduler.

A larger held-out evaluation and more detail on predictor training, rollout collection, and offline costs would also strengthen the paper.

I am keeping my rating at 2 (Weak rejection) and my confidence at 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.