acceptodds
Under review as a conference paper at ICLR 2027

Dispersed Any-Order Speculative Decoding for Exact Parallel Autoregressive Image Generation

Abstract

Autoregressive image models generate one token per forward pass, which makes inference slow. Speculative decoding can speed up inference by generating multiple tokens per pass without changing the sampled distribution. The speedup depends on how much the drafter's distribution disagrees with the target's, and we show that this disagreement has two sources: 1) the drafter can be a different model than the target, and 2) it can condition on different tokens than the target since the drafter generates in parallel while the target conditions each draft on the ones before it. Recent self-speculative methods use the target model as its own drafter, which removes the first source. However, for left-to-right generation, since the drafted positions are adjacent and strongly dependent on one another, the second source of disagreement remains large. We find that disagreement decreases as distance between draft tokens increases, reducing the second source of disagreement. Based on this, we propose Dispersed Any-Order Speculative Decoding (DSD), which uses an any-order model as its own drafter and verifier along an order whose consecutive positions are far apart. DSD requires no additional training and samples from the exact target model's distribution. On three any-order image models, DSD samples the exact distribution with up to fewer forward passes than sequential decoding, fewer than the best existing self-drafting method, and fewer than speculative decoding with a trained drafter, and it matches the quality of the models' lossy sampling with – fewer steps.

Primary area:
generative models
Keywords:
speculative decoding, autoregressive image models, any-order image models, parallel decoding

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_vzQsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

I understand this paper as two points 1 any order AR model provides conditional probability with respect to tokens at arbitrary locations, while fixed design may not have, e.g., NAR, AR, PAR 2 the dependency between tokens gradually decreases as they are far away. This allows using any-order AR as a powerful drafter and verifier, letting the parallel-generated token far away be less entangled. This formulation is good.

1 As the author claimed in Limitation, this method can only apply in any-order AR, making the scope something small. For example, currently it still lacks t2i any-order AR model, making the application limited

2 The observation of "the dependency between tokens gradually decreases as they are far away" is not novel; please cite LPD and NAR at the proper location.

[1] Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation

[2] Neighboring autoregressive modeling for efficient visual generation

3 The reported cost metric (target forward passes per image) flatters the method, since drafting and verification are fused into one heavier pass while wall-clock is given only at batch size 1 and γ is tuned favorably;

Critical strengths

1 The observation well motivated the method.

  1. The method is simple and effective. The result is promising and well-verified.
Critical weaknesses

1 As the author claimed in Limitation, this method can only apply in any-order AR, making the scope something small. For example, currently it still lacks t2i any-order AR model, making the application limited

2 The observation of "the dependency between tokens gradually decreases as they are far away" is not novel; please cite LPD and NAR at the proper location.

[1] Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation

[2] Neighboring autoregressive modeling for efficient visual generation

3 The reported cost metric (target forward passes per image) flatters the method, since drafting and verification are fused into one heavier pass while wall-clock is given only at batch size 1 and γ is tuned favorably; a total-FLOPs or latency-vs-batch-size curve would be fairer. The lossy comparison is also framed advantageously: invoking KL to claim exactness is circular (guaranteed by construction), and switching to FDr⁶ inflates the gap to 5.2–10.1×, whereas under FID the practical gap is closer to 3.5×.

I am willing to raise my score if the authors address these concerns

Questions for authors

Please refer to the above

4 comments

We thank the reviewer for the very useful feedback and for noting that the method is well motivated, simple and effective, with well-verified results. We address each concern below, with new results on a text-to-image model and a detailed throughput analysis, and we hope that these resolve them.

W1 - Text-to-image model

Any-order autoregressive image models are relatively recent, but that line of work is growing fast (RandAR at CVPR 2025, ARPG and LPD at ICLR 2026) and is gaining attention for zero-shot properties that raster models lack (e.g., parallel decoding, inpainting, outpainting). To complement our results, we also ran DSD on an any-order text-to-image model, RLlamaGen [1]. We report the mean target pass count per image on COCO captions:

samplertarget passesreduction against sequential
SJD, raster153.01.7x
SCD, raster116.82.2x
DSD34.17.5x

DSD needs 3.4x fewer passes than standard SCD along raster, which aligns with the class-conditional models. We agree that DSD is limited to any-order models, as stated in our limitations. We will add the text-to-image results to the paper.

[1] NEP: Autoregressive Image Editing via Next Editing Token Prediction

W2 - Citing LPD and NAR at relevant locations

We agree and thank the reviewer for the references. The paper discusses LPD in the related work (Section 6) but does not cite it (or make a connection with it) where the “locality vs. dependence” observation is introduced. NAR, on the other hand, is missing. We will cite both in Section 3. LPD and NAR use locality to design lossy parallel decoders. In our work it is one part of a broader analysis of speculative decoding, where we separate the disagreement between drafter and target into a part due to the model and another part due to the conditioning. Drafting with the target model itself removes the first, and the remaining disagreement is caused by the dependence between tokens drafted together, which can be reduced with a dispersed order. We show that an any-order model can achieve this without any additional training.

W3 - Cost metric and the lossy comparison

We agree that the pass count alone does not give the full picture in terms of cost. We would first like to clarify that, by default DSD adapts SCD, where a pass takes the γ drafts (current window) as input and returns the model's distribution at each of these positions. These distributions are used to verify the current drafts and to draw the drafts of the next step. So, a pass evaluates at most γ positions and we do not perform a separate pass for drafting. The pass therefore has the same computational load as the standard SCD and SJD baselines, which all have the lightest forward passes (lowest computational load) among the speculative samplers we report. For comparison, fused drafting evaluates γ(γ+1) positions per pass (since verification and drafting are fused by predrafting for all potential verification outcomes) and the trained drafters add drafter passes. The target pass count therefore does not favor DSD against any of the speculative decoding baselines. Against sequential decoding, a DSD pass costs 1.2x a sequential pass at batch size 1. We will make the cost clearer in Section 5.

As suggested, we also measured the throughput vs. batch size with the same settings (one H100) on the largest checkpoint of each family (where a pass costs the most). The optimal draft window depends on the batch size (smaller window for larger batch size), so, for each batch size, we report the throughput (images per second) for each speculative sampler at its fastest window (among 8, 16, 32 and 64):

modelsamplerB = 14163264
RandAR-0.7BDSD1.474.418.659.7110.3
SCD, raster0.431.353.574.915.28
sequential0.160.642.575.109.50
ARPG-XXLDSD1.194.6115.324.831.6
SCD, raster0.381.434.658.3813.4
sequential0.150.612.454.929.24
LPD-XXLDSD1.906.1711.814.115.2
SCD, raster0.712.274.696.376.56
sequential0.220.883.517.0614.0

DSD is ~2x to 3.4x faster than standard SCD along raster at every batch size. It also stays ahead of sequential decoding at every batch size (up to 64). Speculative decoding trades computation (FLOPs) for latency, so naturally, this advantage is largest at small batch sizes and shrinks as the batch grows (as we become compute-bound). DSD also needs less FLOPs since, against the other speculative samplers at matched draft window size, it needs fewer passes. We will add the throughput curves for every window to the paper.

Regarding the KL test, we agree that exactness holds by construction. We only use it to measure how far the lossy samplers are from the model's distribution at the step counts where FID no longer distinguishes between them. We will make this explicit in the paper.

Regarding FID and FDr^6, they both are Fréchet distances and differ only in the representations used to compute the distance. FID uses the Inception representation only, while FDr^6 averages over six representations (Inception, ConvNeXt, MAE, DINOv2, SigLIP and CLIP), including ones from more recent models. We understand that FID is still widely used and considered the standard, which is why the paper reports both. We will give both speedup ranges in the abstract to make this clearer.

Full throughput table:

To complement our response to W3, here is the throughput for every draft window and batch size (same setup as in the previous table) where each cell reports DSD / SCD:

modelwindowB = 14163264
RandAR-0.7B80.62 / 0.342.08 / 1.125.55 / 3.057.64 / 4.488.69 / 5.28
160.98 / 0.413.21 / 1.357.56 / 3.579.71 / 4.9110.3 / 5.07
321.20 / 0.414.22 / 1.118.65 / 3.449.50 / 3.759.35 / 3.93
641.47 / 0.434.41 / 1.317.34 / 2.467.47 / 2.557.41 / 2.56
sequential0.160.642.575.109.50
ARPG-XXL80.58 / 0.312.21 / 1.118.94 / 4.4716.7 / 8.3127.6 / 13.4
160.92 / 0.353.46 / 1.1713.0 / 4.6524.8 / 8.3831.6 / 10.3
321.16 / 0.384.12 / 1.4315.3 / 4.6023.4 / 6.7024.8 / 7.49
641.19 / 0.364.61 / 1.4013.2 / 4.0714.5 / 4.3614.0 / 3.96
sequential0.150.612.454.929.24
LPD-XXL80.89 / 0.513.02 / 1.688.20 / 4.3610.7 / 5.8512.1 / 5.27
161.34 / 0.664.37 / 1.9911.8 / 4.6914.1 / 6.3715.2 / 6.56
321.88 / 0.705.45 / 2.2611.7 / 3.8412.6 / 4.7914.5 / 5.24
641.90 / 0.716.17 / 2.279.12 / 2.8110.6 / 3.2310.4 / 3.41
sequential0.220.883.517.0614.0

Thank you for the response. I have no more concerns and raise my score.

Reviewer_JLFmOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Dispersed Any-Order Speculative Decoding (DSD), a training-free acceleration method for any-order autoregressive image models. The central observation is that self-drafting removes model mismatch, leaving disagreement caused by differences in conditioning between parallel drafting and sequential verification. Since spatially distant image tokens tend to depend less on one another, DSD uses a dispersed bit-reversal generation order to improve draft acceptance. The method preserves the sequential distribution associated with the chosen order. Experiments on seven checkpoints from three ImageNet class-conditional model families show substantial reductions in target forward passes. Appendix D also reports 7.6–27× latency improvements over sequential decoding at batch size one.

Critical strengths
  1. A simple and useful connection between spatial dependence and speculative acceptance. The paper makes generation order an explicit design variable for self-speculative decoding. Figures 2–3 provide a coherent empirical explanation: using the target model as its own drafter eliminates model mismatch, while dispersing drafted positions reduces conditioning mismatch. This gives a practical way to improve acceptance without training an additional model.

  2. Large and consistent gains across model families and existing samplers. Table 2 is particularly informative because changing the order improves all three evaluated self-speculative samplers. The gains therefore extend beyond one particular sampler. The evaluation includes multiple model sizes and a 1,024-token setting, and Appendix D confirms that the reduction in passes translates into substantial latency improvements over sequential decoding.

Critical weaknesses
  1. The order ablations do not fully establish the contribution of the proposed schedule or clarify comparisons across target distributions. Random ordering appears in the acceptance analysis, but the main evaluation does not compare complete DSD runs under random ordering, bit-reversal ordering, and the dispersed schedules used by related image decoders. Consequently, it is difficult to determine how much benefit specifically comes from bit reversal beyond using a broadly dispersed order. Moreover, exactness is with respect to the sequential distribution under the chosen generation order. Changing that order can change the joint distribution defined by a learned any-order model. The paper acknowledges this limitation, but the main comparisons would be stronger with explicit sequential quality results for each evaluated order.

  2. The implemented sampler and its computational advantage need more direct documentation. Algorithm 1 describes separate drafting and verification, whereas the principal results use SCD and claim one target pass per step. The paper cites the relevant techniques, but does not sufficiently explain the actual masks, cached states, and proposal distributions used when combining them with DSD. Appendix D establishes latency gains over sequential and lossy sampling, but does not report latency against SCD, SJD, fused drafting, or trained drafters. Since these methods can have different costs per pass, the forward-pass comparisons alone do not establish the practical speedup over the strongest speculative baseline.

Questions for authors
  1. Could you report target passes and sample quality for the same self-speculative sampler under random ordering, bit-reversal ordering, and an existing dispersed image-generation schedule? Please also identify the sequential order used for the quality reference in Figures 6 and 9.

  2. Could you explain how the SCD implementation realizes DSD, including which proposal distribution is used for acceptance and residual resampling, and how drafting and verification share a forward pass while preserving exactness?

  3. Could you provide batch-size-one latency and peak memory comparisons against raster-order SCD and fused drafting under the same hardware and implementation settings? These measurements would directly substantiate the practical advantage over existing exact speculative samplers.

4 comments

We thank the reviewer for the thorough review and for highlighting the simple and useful insight behind the method and its large consistent gains across models and samplers. The three questions were very helpful and led to clarifications that we will add to the paper.

W1 / Q1 - Order ablation

We ran SCD (draft window 64) along orders of increasing dispersion: raster (as in standard SCD) and four orders made possible by adapting SCD to an any-order model, including random orders and Halton [1]. We report the dispersion (mean distance in grid between a position and the closest of its neighbors in the order) and number of target passes per image (1000 images):

orderdispersionRandAR-0.7BARPG-LLPD-L
raster1.191.076.665.5
raster with stride 22.146.436.732.2
random3.429.525.723.7
Halton4.125.622.320.7
bit-reversal4.525.722.521.1

SJD shows the same pattern (under the same settings):

orderRandAR-0.7BARPG-LLPD-L
random43.739.937.4
Halton35.431.829.9
bit-reversal35.231.729.8

On every checkpoint and sampler we tested, the speedup follows the dispersion of the order up to a certain point. It grows quickly from raster to random orders (which are already fairly dispersed) and levels off among the most dispersed orders. We chose bit-reversal for the paper because it is the natural order for DSD. Since we cannot know where every drafting step will start (since it depends on the verification outcome), every window of consecutive positions has to be dispersed, and bit-reversal ensures this by construction. As stated in Section 6, DSD works with any dispersed order.

Sample quality (FID 50k):

orderRandAR-0.7BARPG-LLPD-L
random2.132.652.60
Halton2.182.932.72
bit-reversal2.082.682.62

The FID for bit-reversal, Halton, and random (which the models are trained on) are within the same range. In Figures 6 and 9, the lossy and sequential baselines all run along the same dispersed order as DSD.

[1] Halton Scheduler For Masked Generative Image Transformer

W2 / Q2 - How DSD is implemented with SCD

We agree that the paper should make each sampler’s implementation clearer. We will add a description and algorithm box for each sampler we’ve adapted. For SCD’s integration with DSD, let x_i be the token of the i-th position of a dispersed order. At step t, the first c tokens are committed, and each of the next W positions (the draft window) has a draft x_i that was sampled from a distribution q_i^(t). A single target pass verifies these drafts and also produces the drafts of step t+1. The process goes as follows:

  1. The target model evaluates the W drafts in one pass. Its input is the committed tokens, whose KVs are cached, followed by the position query and the draft of each position of the window. At the query of position i it outputs p_i^(t), the distribution of x_i given the committed tokens and the drafts before position i
  2. We verify the drafts in order with the standard speculative decoding rule, where x_i is accepted with probability min(1, p_i^(t)(x_i) / q_i^(t)(x_i)). Let r be the position of the first rejected token, the drafts before r are committed, and x_r is resampled from the residual norm(max(0, p_r^(t) - q_r^(t))) and then committed as well
  3. Unlike standard speculative decoding methods, the drafts after r are not discarded. Each one is kept with probability min(1, p_i^(t)(x_i) / q_i^(t)(x_i)) and otherwise resampled from the residual (the coupling of SCD). In both cases x_i is now an exact sample of p_i^(t), so its proposal distribution at the next step is q_i^(t+1) = p_i^(t)
  4. The positions committed in step 2 leave the window, and the next positions of the order take their place. The window of step t+1 thus consists of the drafts of step 3 followed by these new positions. Those new positions start from a uniformly random draft

So, the pass at step t verifies the drafts of step t (p_i^(t) against q_i^(t), where q_i^(t)=p_i^(t-1) for the drafts already in the window), and gives the proposal distributions of step t+1 (q_i^(t+1) = p_i^(t)). Algorithm 1 shows drafting and verification as two passes only for generality (to apply to all samplers).

DSD adapts self-speculative samplers such as SCD to any-order models, which gives control over the positions that are drafted together. The drafts of a window are produced in parallel, each conditioned on other drafts that have not been verified yet. With a raster model there is no such control since the model can only predict the token at the next adjacent position. Adjacent tokens depend strongly on each other, so when one of the draft tokens is rejected the drafts after it are no longer valid. An any-order model can predict any position, so DSD drafts positions that are far apart. Since dependence falls with distance, each draft is then determined almost entirely by the tokens already committed (without having to rely on the draft tokens). It is close to a sample of the model’s distribution given the committed tokens, and it remains valid whether or not the other drafts are accepted, so many more drafts are committed per pass.

W2 / Q3 - Cost per pass, latency and peak memory vs. the other exact samplers

We agree that pass counts alone do not establish the speedup when passes differ in cost. A pass costs the same for DSD, SCD and SJD, so for these samplers the reduction in passes translates to reduction in latency. The two exceptions are fused drafting (whose passes are heavier because each one also evaluates the drafts of every potential acceptance outcome) and the trained drafters (which run drafter passes in addition to the target passes). For both, the pass counts of Tables 1 and 2 leave out part of the cost, so they understate the speedup of DSD. We measured the self-speculative samplers with the same settings (one H100, batch size 1, draft window 64). Here’s a table showing the latency (seconds per image) and peak memory:

modelDSDSCD (raster)SJD (raster)fused drafting (raster)sequential
RandAR-0.7B0.66 s, 3.4 GB2.33 s, 3.4 GB3.34 s, 3.4 GB42.1 s, 7.1 GB5.63 s, 3.3 GB
ARPG-L0.30 s, 1.4 GB0.99 s, 1.4 GB1.59 s, 1.4 GB2.30 s, 2.7 GB3.01 s, 1.4 GB
LPD-L0.17 s, 1.5 GB0.54 s, 1.5 GB0.93 s, 1.5 GB12.9 s, 4.1 GB2.31 s, 1.4 GB

DSD is 3.2x to 3.5x faster than standard SCD along raster and 5.1x to 5.4x faster than standard SJD along raster, with the same peak memory. Fused drafting has costlier passes because it evaluates the drafts of every acceptance outcome in the same pass (which grows substantially with γ=64). At its fastest draft window (8 or 16), it takes 2.75 s, 1.19 s and 0.83 s on the three models, which is still 4.0x to 4.9x slower than DSD.

For the trained drafters, the target passes alone give a lower bound on latency, since counting the drafter passes would only add to the cost. The best trained drafter (DFlash) then needs at least 3.1 s per image on RandAR-0.7B and 1.4 s on ARPG-L before any drafter time is counted, so DSD is at least 4.5x faster. We will add these measurements to the paper and state in Tables 1 and 2 which costs are excluded from the pass count.

Thank you for your responses and I think it resolves my concerns.

Comments on the paper

4 comments

We thank the reviewer for the detailed review and appreciate the positive comments about our analysis and the consistency of our results. We answer each question below, with new results that we will add to the paper.

W1 / Q1 - Experimental coverage

On the six 256-token checkpoints (the smallest and the largest released model of RandAR, ARPG, and LPD), DSD needs 21 to 27 passes, 9.5x to 12.2x fewer than sequential decoding and 3.0x to 3.5x fewer than the best existing exact sampler. The gains therefore hold without the 1024-token model, and we will state the corresponding range for each sequence length (256 and 1024 tokens) in the abstract.

To the best of our knowledge, the longest sequence length for released any-order image models is 1024 tokens (LPD at 512x512). We are not aware of one at 2048 tokens. We ran DSD on LPD’s second released 1024-token checkpoint, LPD-XL 512: DSD needs 35 passes, 29.2x fewer than sequential decoding and 4.6x fewer than standard SCD along raster (162.7 passes), which is similar to the speedup on LPD-L 512 (29.8x fewer passes than sequential decoding). In the paper, the reduction on LPD-L is larger at 1024 tokens than at 256 tokens (29.8x vs. 12.2x fewer passes than sequential decoding), and LPD-XL 512 corroborates the 1024-token result. We do not present this as conclusive evidence that the gain increases with sequence length. Our main result is that the gain holds on every checkpoint we tested. We will reword the sentence “the gain grows with the number of tokens” in Section 5.2 accordingly.

We also ran experiments on an any-order text-to-image model, RLlamaGen [1]. On 256 COCO captions, DSD needs 34.1 passes (~7.5x fewer than sequential decoding), while SCD and SJD along raster need 116.8 and 153.0.

Bit-reversal is compatible with any grid size (does not need to be square). The order is computed from the row index and the column index of each position, and the two are handled separately, so the number of rows and the number of columns do not have to match and can be of any size.

[1] NEP: Autoregressive Image Editing via Next Editing Token Prediction

W2 / Q2 - Comparison with trained drafters

We would like to clarify that the main metric we use for cost (number of target passes) actually favors the trained-drafter baselines rather than our method (DSD). For DSD, since drafting and verification use the target model, all passes are included in the “target passes” metric. On the other hand, trained-drafter methods like EAGLE and DFlash use a separate drafter: EAGLE runs one drafter pass per token in the draft (autoregressively), and DFlash runs one drafter pass to output all the draft tokens for a step in parallel. To keep a single, simple cost metric for all methods, we only count the target passes (i.e., verification passes) for the trained-drafter baselines and leave out the drafter passes. Counting them (or their wall-clock time) can only increase the speedup of DSD over these baselines. Even with this advantage, the trained-drafter baselines significantly underperform DSD. For example, on an H100, the target passes alone of the best trained drafter (DFlash) take at least 3.1 s per image on RandAR-0.7B and 1.4 s on ARPG-L, against 0.66 s and 0.30 s for DSD. Regarding training budget, DSD does not need any additional training, so there is no matched-budget trained drafter to compare with. If the reviewer has a different comparison in mind, we would be glad to provide it.

Q3 - How does the chosen order compare to raster in terms of quality?

For an any-order model that fitted the data perfectly, every order would have the same distribution. However, in practice the distribution and sample quality depend on the order. On RandAR-0.7B the likelihood test of Table 5 separates samples along bit-reversal from samples along raster (AUC 1). The following table shows the negative log-likelihood (2000 validation ImageNet images) and the FID (50k images):

modelordernegative log-likelihoodFID
RandAR-0.7Braster1938.95.94
RandAR-0.7Brandom1915.22.13
RandAR-0.7Bbit-reversal1916.02.08
LPD-Lraster1941.13.22
LPD-Lrandom1928.42.60
LPD-Lbit-reversal1927.92.62
ARPG-Lraster1962.410.76
ARPG-Lrandom1932.02.65
ARPG-Lbit-reversal1939.22.68

The effect of order on quality depends on the checkpoint. Unlike raster, bit-reversal is generally close to random orders (which the models were trained on).

Q4 - What does the pass count variance across images look like?

On ARPG-L, bit-reversal needs 22.5 +/- 3.1 passes (min: 15, max: 36), while raster needs 76.6 +/- 15.2 (min: 44, max: 130). The worst image along bit-reversal needs fewer passes than the best image along raster. We will add the standard deviation to Figure 5(c).

Q5 - How many draft tokens get rejected per step?

The number of drafts committed per pass grows with the draft window. On ARPG-L it is 4.3 at a window of 8, 9.7 at 32 and 11.7 at 64 (Figure 5(a)). As is standard in speculative decoding, we can spend additional parallel computation to reduce the number of sequential passes (latency), and the draft window should be chosen for the hardware and the expected batch size. At batch size 1, a pass with 64 drafts costs only 1.2x a sequential pass in terms of latency, so a large window gives the lowest latency, whereas, at a large batch size, a smaller window gives the highest throughput. We will include a throughput analysis in the paper. With SCD, our default, the drafts after a rejection are not discarded; they stay in the window and are kept or resampled according to their new conditionals. Block verification is compatible with our method; we will mention it as an extension.

Q6 - Could you report the KL analysis (Appendix C) at matched-FID step counts rather than at 64?

We ran the analysis on the three checkpoints of the lossy comparison (the step counts where FID ties are in bold): KL(lossy || model) in nats per image and the AUC of the likelihood test between lossy and exact samples:

modelstepsKL (nats per image)AUC
RandAR-0.7B6416.4 +/- 0.31.000
RandAR-0.7B8811.0 +/- 0.20.998
RandAR-0.7B1285.5 +/- 0.10.983
ARPG-XXL3259.0 +/- 0.71.000
ARPG-XXL6428.3 +/- 0.41.000
ARPG-XXL12813.0 +/- 0.21.000
LPD-XXL3246.1 +/- 0.81.000
LPD-XXL6417.9 +/- 0.40.999
LPD-XXL1282.8 +/- 0.10.934

At the step counts where FID ties, the lossy samples are still 11 to 46 nats per image away from the model’s distribution. The likelihood test can tell them apart from exact samples almost perfectly (AUC >= 0.998). We will replace the table of Appendix C with this one.

Thanks for the rebuttal. It has addressed my concerns. I would keep my rating as positive side.

AuthorsOpenReview ↗

Revised paper

Dear Area Chair and Reviewers,

We sincerely thank the reviewers for their constructive feedback. All three reviewers have confirmed that our responses resolved their concerns, and we have incorporated the results and clarifications from the discussion into the revised paper (changes in blue), including:

  • Results for new models: LPD-XL 512 and the text-to-image model RLlamaGen (Table 1, Section 5.1)
  • Latency and throughput against the other exact samplers (Section 5.2, Appendix D)
  • Order ablation (Appendix B)
  • DSD with SCD (Algorithm 2, Appendix A) and the cost each pass count includes (Sections 4.1 and 5.1, Table 1)
  • KL test at the matched-FID step counts (Appendix C)
  • Speedup ranges for each sequence length and on both FID and FDr^6 (Abstract)
  • Citations of LPD and NAR (Section 3), block verification (Section 6), and the per-image standard deviation (Figure 5c)

We thank the reviewers again for the discussion, which strengthened the paper, and the Area Chair for their time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.