acceptodds
Under review as a conference paper at ICLR 2027

SSD: Spatially Speculative Decoding Accelerates Autoregressive Image Generation

Abstract

Autoregressive image models treat images as 1D token sequences, inheriting the next-token factorization of language models. This flattening discards a useful property of images: visual tokens are organized and correlated in two-dimensional space. We introduce Spatially Speculative Decoding (SSD), an inference-time decoding framework that exploits this spatial structure for accelerated generation. Rather than speculating only along the flattened sequence, SSD predicts both the adjacent horizontal token and the token directly below it, allowing multiple spatial directions to advance in parallel. This reduces the number of backbone forward evaluations and alleviates the memory bottleneck of autoregressive decoding. SSD accelerates image generation by up to 11.03× in wall-clock time while maintaining generation quality on DPG-Bench and GenEval. These results show that spatial structure provides a simple and effective source of parallelism for faster autoregressive image generation.

Primary area:
generative models
Keywords:
Autoregressive Image Generation, Multi-Token Prediction and Speculative Decoding.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_297YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Spatially Speculative Decoding (SSD), an inference acceleration framework for autoregressive image generation that exploits the two-dimensional spatial correlations of visual tokens. SSD introduces lightweight horizontal and vertical multi-token prediction (MTP) heads to predict future hidden states while keeping the pretrained backbone frozen. By leveraging vertical prediction, SSD enables row-level parallel drafting, followed by iterative verification and correction to maintain generation quality. Experiments on Janus-Pro-7B, Lumina-mGPT-7B, and Emu3-8B demonstrate wall-clock speedups of 5.70×, 7.35×, and 11.03×, respectively, with relatively small degradation in generation quality.

Critical strengths

1.The motivation is well supported by the analysis of spatial correlations in visual tokens. The demonstrated feasibility of vertical prediction naturally leads to the proposed row-level parallel decoding approach.

  1. The experiments cover multiple backbone models, competitive baselines, and ablation studies. The reported results demonstrate substantial decoding speedups with relatively minor degradation in generation quality.

  2. The paper is well organized and clearly written, with a coherent progression from motivation and analysis to methodology and experimental validation.

Critical weaknesses
  1. Table 1 shows unexplained discrepancies between the step and wall-clock speedups of SJD and GSD across different backbones, particularly on Lumina-mGPT and Emu3-8B. For example, on Emu3-8B, SJD achieves only 1.08× wall-clock speedup despite a 2.24× step reduction, whereas GSD achieves 2.90× speedup with a 3.46× step reduction. Given that GSD largely follows the same Jacobi decoding workflow as SJD while introducing additional sorting and grouping overhead, these differences in computational efficiency are difficult to explain. Even these relatively peripheral baselines show unexplained inconsistencies. This raises my concerns about the reliability of the more critical SSD speedup results.

  2. The reported training costs do not account for the computational expense of generating training data and extracting backbone hidden states.

My primary concern is Weakness 1, particularly the unexplained discrepancies between step reduction and wall-clock speedups across different baselines, which raise questions about the reliability of the reported SSD speedup results. The incomplete accounting of training costs in Weakness 2 is a secondary concern. I encourage the authors to focus their rebuttal on explaining the discrepancies identified in Weakness 1.

If this concern can be satisfactorily addressed, I would be inclined to support acceptance of this paper.

Questions for authors
  1. Could the authors explain the discrepancies between SJD and GSD across all evaluated backbones, particularly on Emu3-8B, at the decoding-workflow and computational-cost levels? An explanation based solely on improved acceptance rates is insufficient to account for the reported differences between step and wall-clock speedups.

  2. Could the authors report the total computational cost, including autoregressive data generation, hidden-state extraction, and MTP head training?

6 comments

We thank the reviewer for recognizing SSD's strong empirical motivation, substantial speedups, and clear and coherent presentation.

Weakness 1: Clarification of implementation details: discrepancies in speedups between SJD and GSD

Our reported SSD speedups are reliable and based on a fair comparison with AR. We thank the reviewer for pointing out the discrepancy between SJD [1] and GSD [2]. After inspecting GSD's code, we find that GSD's implementation batches the conditional and unconditional CFG branches, whereas SJD and AR use serial CFG on Lumina-mGPT and Emu3. Batching runs both branches in one backbone call rather than two serial calls, reducing latency without changing step counts. SSD uses the same serial CFG strategy as AR in Table 1, ensuring a fair comparison.

We therefore evaluate SJD, GSD, and SSD under both serial and batched CFG strategies on the same hardware using 100 DPG-Bench prompts. See the results in the tables below; speedups are relative to the same serial AR baseline.

Batched CFG.

BackboneMethodStep speedupLatency (s)Latency speedup
Lumina-mGPT-7BSJD2.23×2.23\times32.902.92×2.92\times
Lumina-mGPT-7BGSD3.24×3.24\times24.243.97×3.97\times
Lumina-mGPT-7BSSD8.40×8.40\times7.5312.78×12.78\times
Emu3-8BSJD2.25×2.25\times171.751.64×1.64\times
Emu3-8BGSD3.27×3.27\times120.912.33×2.33\times
Emu3-8BSSD14.32×14.32\times19.1014.76×14.76\times

Serial CFG.

BackboneMethodStep speedupLatency (s)Latency speedup
Lumina-mGPT-7BSJD2.23×2.23\times59.781.61×1.61\times
Lumina-mGPT-7BGSD3.24×3.24\times42.862.25×2.25\times
Lumina-mGPT-7BSSD8.40×8.40\times13.157.32×7.32\times
Emu3-8BSJD2.25×2.25\times260.441.08×1.08\times
Emu3-8BGSD3.27×3.27\times182.311.55×1.55\times
Emu3-8BSSD14.32×14.32\times25.1011.23×11.23\times

Under matched CFG strategies, SJD and GSD exhibit consistent overall latency trends: batching reduces latency for both methods while leaving step counts essentially unchanged. SSD outperforms both methods under matched experimental settings. We will make these CFG settings explicit in the revision.

Weakness 2: Additional reporting of computational costs

We thank the reviewer for bringing this up. We report the computational costs of generating training data and extracting backbone hidden states below. We will include these costs in the revision.

  1. Training-data generation is a separate, one-time cost. See the costs in the table below. We will report data generation separately from training in the revision.
ModelData generation (GPU-h)Training, including hidden-state extraction (GPU-h)Total (GPU-h)
Janus-Pro-7B29.73.0232.7
Lumina-mGPT-7B27.785.5533.33
Emu3-8B40.973.3144.28
  1. Hidden-state extraction is already included in training. Hidden states are computed by the frozen backbone within each training step. The reported training costs of 3.02, 5.55, and 3.31 GPU-hours for Janus-Pro, Lumina-mGPT, and Emu3 therefore include both hidden-state extraction and head optimization.

To be continued.

Question 1: Clarification of the discrepancies between SJD and GSD

We thank the reviewer for pointing this out. After inspecting GSD's code [2], we find that GSD's implementation batches the conditional and unconditional CFG branches, whereas SJD [1] and AR use serial CFG on Lumina-mGPT and Emu3. Batching runs both branches in one backbone call rather than two serial calls, reducing latency without changing step counts. SSD uses the same serial CFG strategy as AR in Table 1, ensuring a fair comparison.

We therefore evaluate SJD, GSD, and SSD under both serial and batched CFG strategies on the same hardware using 100 DPG-Bench prompts. See the results in the tables below; speedups are relative to the same serial AR baseline.

Batched CFG.

BackboneMethodStep speedupLatency (s)Latency speedup
Lumina-mGPT-7BSJD2.23×2.23\times32.902.92×2.92\times
Lumina-mGPT-7BGSD3.24×3.24\times24.243.97×3.97\times
Lumina-mGPT-7BSSD8.40×8.40\times7.5312.78×12.78\times
Emu3-8BSJD2.25×2.25\times171.751.64×1.64\times
Emu3-8BGSD3.27×3.27\times120.912.33×2.33\times
Emu3-8BSSD14.32×14.32\times19.1014.76×14.76\times

Serial CFG.

BackboneMethodStep speedupLatency (s)Latency speedup
Lumina-mGPT-7BSJD2.23×2.23\times59.781.61×1.61\times
Lumina-mGPT-7BGSD3.24×3.24\times42.862.25×2.25\times
Lumina-mGPT-7BSSD8.40×8.40\times13.157.32×7.32\times
Emu3-8BSJD2.25×2.25\times260.441.08×1.08\times
Emu3-8BGSD3.27×3.27\times182.311.55×1.55\times
Emu3-8BSSD14.32×14.32\times25.1011.23×11.23\times

Under matched CFG strategies, SJD and GSD exhibit consistent overall latency trends: batching reduces latency for both methods while leaving step counts essentially unchanged. SSD outperforms both methods under matched experimental settings. We will make these CFG settings explicit in the revision.

Question 2: Additional reporting of the total computational cost

We thank the reviewer for this suggestion. We report the total computational cost in the table below.

  1. Autoregressive data generation: Data generation is a separate, one-time cost. The costs are 29.7 GPU-hours for Janus-Pro, 27.78 for Lumina-mGPT, and 40.97 for Emu3.

  2. Hidden-state extraction and MTP head training: Hidden-state extraction is already included in MTP head training. Hidden states are computed by the frozen backbone within each training step. The reported training costs of 3.02, 5.55, and 3.31 GPU-hours for Janus-Pro, Lumina-mGPT, and Emu3 therefore include both hidden-state extraction and head optimization.

BackboneAR data generation (GPU-h)Head training, including hidden-state extraction (GPU-h)Total (GPU-h)
Janus-Pro-7B29.73.0232.7
Lumina-mGPT-7B27.785.5533.33
Emu3-8B40.973.3144.28

References

[1] Teng et al. Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding. ICLR, 2025.

[2] So et al. Grouped Speculative Decoding for Autoregressive Image Generation. ICCV, 2025.

AuthorsOpenReview ↗

We welcome your feedback.

We thank the reviewer for their time and consideration. We hope our responses have addressed their questions and concerns, and would appreciate any further feedback. We would be happy to clarify any remaining points.

Reviewer_K5fmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spatially Speculative Decoding (SSD) to accelerate autoregressive image generation by exploiting the two-dimensional spatial structure of visual tokens. Keeping the pretrained backbone frozen, SSD trains auxiliary prediction heads to draft tokens along horizontal and vertical directions, enabling parallel drafting of entire rows. The backbone then refines the drafted blocks through a fixed number of parallel verification-and-correction rounds, reducing sequential forward evaluations. Experiments evaluate SSD on Janus-Pro-7B, Lumina-mGPT-7B, and Emu3-8B using DPG-Bench and GenEval, with additional perceptual quality and human preference evaluations on Janus-Pro.

Critical strengths
  1. Substantial measured acceleration with a competitive speed–quality trade-off. Relative to the fastest evaluated baseline, SSD is approximately 2.2–3.1× faster while closely matching DPG-Bench scores on Janus-Pro and Lumina-mGPT and improving them on Emu3.
  2. Broad applicability. SSD accelerates three models with different tokenizers and sequence lengths while keeping their backbones frozen, supporting its practicality for existing autoregressive image generators.
Critical weaknesses
  1. Missing comparisons. Although the paper discusses PJD and MuLo-SD, neither is included in the experiments. PJD already provides training-free, two-dimensional parallel refinement, making it particularly relevant. It remains unclear how much SSD improves over existing spatial decoding approaches.
  2. Insufficient evidence for quality preservation. At the reported operating points, GenEval scores decrease from 0.77 to 0.73 on Janus-Pro, 0.50 to 0.46 on Lumina-mGPT, and 0.45 to 0.40 on Emu3. Additional perceptual and human evaluations are limited to Janus-Pro. These results leave uncertainty about whether the headline speedups consistently preserve compositional generation quality across backbones.
Questions for authors
  1. Empirical clarification. How does SSD compare with PJD under the same backbone, hardware, and matched generation quality?
  2. Stronger evaluation. How do SSD’s latency and throughput gains vary with inference batch size? Does its advantage persist at larger batches, where the AR baseline may already achieve higher GPU utilization?
4 comments

We thank the reviewer for recognizing SSD's substantial acceleration and cross-model compatibility.

Weakness 1: Additional comparisons with PJD and MuLo-SD

We thank the reviewer for this suggestion. We conducted additional experiments with PJD [1] and MuLo-SD [2]. SSD is faster than PJD and MuLo-SD, with comparable DPG-Bench quality. See the table below. We will cite both methods and include the new results in the revision.

BackboneMethodLatency (s, ↓\downarrow)Latency speedup vs. AR (↑\uparrow)DPG-Bench (↑\uparrow)
Janus-Pro-7BAR7.871.00×1.00\times84.23
Janus-Pro-7BPJD4.231.86×1.86\times83.78
Janus-Pro-7BMuLo-SDN/AN/AN/A
Janus-Pro-7BSSD1.385.70×5.70\times83.40
Lumina-mGPT-7BAR96.231.00×1.00\times76.30
Lumina-mGPT-7BPJD19.005.07×5.07\times74.77
Lumina-mGPT-7BMuLo-SD26.543.63×3.63\times75.15
Lumina-mGPT-7BSSD13.097.35×7.35\times74.57
Emu3-8BAR281.871.00×1.00\times78.69
Emu3-8BPJD54.875.14×5.14\times77.19
Emu3-8BMuLo-SDN/AN/AN/A
Emu3-8BSSD25.5511.03×11.03\times76.03

MuLo-SD's code has not been released, so we reimplemented the method based on the paper [2]. It requires a low-resolution drafter, which limits its applicability and prevents us from using it with Janus-Pro and Emu3 in our setting. Janus-Pro supports only a resolution of 384×384, and Emu3's low-resolution outputs often omit the main subject. We therefore evaluate MuLo-SD only on Lumina-mGPT, using the official 512 checkpoint as the drafter.

Weakness 2: Clarification on quality preservation

SSD explores acceleration without exact distribution preservation, an important line of research that includes ZipAR [3], GSD [4], LANTERN [5], COOL-SD [6], and MuLo-SD [2]. SSD still retains a gap relative to AR's GenEval score. However, SSD outperforms ZipAR [3], the fastest approximate decoding baseline in our paper, while achieving comparable or higher GenEval scores (see the table below). Our approach can be useful when speed is the system's bottleneck and is worth publishing.

ModelLatency (s) ↓\downarrowLatency speedup ↑\uparrowOverall ↑\uparrow
Lumina-mGPT-7B96.231.00×1.00\times0.50
+ ZipAR (w=6w=6)14.556.61×6.61\times0.46
+ SSD (Ours)13.097.35×7.35\times0.46
Janus-Pro-7B7.721.00×1.00\times0.77
+ ZipAR (w=4w=4)1.874.13×4.13\times0.72
+ SSD (Ours)1.455.32×5.32\times0.73
Emu3-8B281.061.00×1.00\times0.45
+ ZipAR (w=4w=4)25.8210.89×10.89\times0.31
+ SSD (Ours)25.0011.24×11.24\times0.40

To be continued.

Question 1: Additional comparison with PJD

We thank the reviewer for this suggestion. We conducted experiments with PJD [1] under the same settings. SSD is faster than PJD, with comparable DPG-Bench quality. See the table below for the results. We will cite PJD and include the new results in the revision.

BackboneMethodLatency (s, ↓\downarrow)Latency speedup vs. AR (↑\uparrow)DPG-Bench (↑\uparrow)
Janus-Pro-7BAR7.871.00×1.00\times84.23
Janus-Pro-7BPJD4.231.86×1.86\times83.78
Janus-Pro-7BSSD1.385.70×5.70\times83.40
Lumina-mGPT-7BAR96.231.00×1.00\times76.30
Lumina-mGPT-7BPJD19.005.07×5.07\times74.77
Lumina-mGPT-7BSSD13.097.35×7.35\times74.57
Emu3-8BAR281.871.00×1.00\times78.69
Emu3-8BPJD54.875.14×5.14\times77.19
Emu3-8BSSD25.5511.03×11.03\times76.03

Question 2: Additional evaluation at different batch sizes

We thank the reviewer for this suggestion. Yes, SSD's advantage persists at larger batch sizes: at a batch size of 16, it is still 2.982.98–3.64×3.64\times faster than AR. Gains generally decrease as the batch size grows, consistent with better GPU utilization by AR.

We conduct batch-size scaling experiments on all three backbones on an RTX PRO 6000 Blackwell GPU, retaining the paper's inference configurations and using the same prompt for both methods. At the same batch size, latency speedup and throughput gain are equal, so each entry reports both.

Batch sizeJanus-Pro-7BLumina-mGPT-7BEmu3-8B
15.16×5.16\times6.64×6.64\times11.02×11.02\times
25.51×5.51\times6.46×6.46\times9.02×9.02\times
44.94×4.94\times4.60×4.60\times5.80×5.80\times
83.09×3.09\times3.45×3.45\times4.99×4.99\times
162.98×2.98\times3.14×3.14\times3.64×3.64\times

At a batch size of 16, throughput increases from 1.10 to 3.28 images/s on Janus-Pro, from 0.091 to 0.286 images/s on Lumina-mGPT, and from 0.029 to 0.105 images/s on Emu3.

References

[1] Liao et al. Parallel Jacobi Decoding for Fast Autoregressive Image Generation. CVPR, 2026.

[2] Peruzzo et al. Multi-Scale Local Speculative Decoding for Image Generation. CVPR, 2026.

[3] He et al. ZipAR: Parallel Autoregressive Image Generation through Spatial Locality. ICML, 2025.

[4] So et al. Grouped Speculative Decoding for Autoregressive Image Generation. ICCV, 2025.

[5] Jang et al. LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding. ICLR, 2025.

[6] Li et al. Annealed Relaxation of Speculative Decoding for Faster Autoregressive Image Generation. AAAI, 2026.

AuthorsOpenReview ↗

We welcome your feedback.

We thank the reviewer for their time and consideration. We hope our responses have addressed their questions and concerns, and would appreciate any further feedback. We would be happy to clarify any remaining points.

AuthorsOpenReview ↗

Brief Follow-Up

We would like to briefly follow up on our responses, as the rebuttal deadline is approaching. We hope they have addressed your concerns and would appreciate any further feedback or questions you may have.

Reviewer_g8VLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Spatially Speculative Decoding (SSD), an approximate decoding framework for accelerating frozen autoregressive image-generation models. SSD trains auxiliary heads to predict future hidden states at horizontal and vertical spatial offsets. Vertical drafting initializes entire rows or multiple rows in parallel, followed by a finite number of verification-and-correction passes. Overall, the results demonstrate substantial latency reductions, accompanied by increased total computation and some degradation in generation metrics.

Critical strengths
  1. Substantial measured acceleration across three backbones. Table 1 reports both absolute latency and reductions in backbone forward evaluations, alongside generation-quality measurements. The reported 5.70×–11.03× speedups substantially exceed those of the evaluated baselines.
  2. A practical design with comparatively transparent reporting of trade-offs. SSD leaves the backbone parameters unchanged and trains only auxiliary predictors. The appendix explicitly acknowledges non-exact sampling and increased total FLOPs and peak memory, while providing additional quality evaluations and verification-budget sweeps. This makes the work a meaningful investigation of latency–quality trade-offs rather than a presentation of speed alone.
Critical weaknesses
  1. The correction rule is insufficiently specified. Section 3.3 combines arg⁡max⁡\arg\max drafting with an acceptance--rejection formula, min⁡(p/q,1)\min(p/q,1) and [p−q]+[p-q]_+, whose standard justification assumes candidates sampled from qq. The reference updates across rounds also require clarification. Although SSD explicitly permits approximate sampling, these ambiguities affect reproducibility and the interpretation of its central correction mechanism.

  2. Quality preservation is not established at the headline speedups. GenEval drops by 0.040.04--0.050.05 across all three models, with larger declines in some compositional categories. Without uncertainty estimates or quality-matched comparisons, the results support a speed--quality trade-off rather than consistently maintained quality. The supplementary FID and human evaluations cover only Janus and do not establish equivalence.

Questions for authors
  1. Correction rule and reference updates. Does the implementation select drafts via t^k=arg⁡max⁡vqk(v)\hat{t}_k = \arg\max_v q_k(v) or sample from qkq_k? Please specify the exact definition of qkq_k in α(t^k)=min⁡(pk(t^k)/qk(t^k),1)\alpha(\hat{t}_k)=\min(p_k(\hat{t}_k)/q_k(\hat{t}_k), 1) and the reference-distribution update rule across rr rounds.

  2. Quality-matched speedups on GenEval. Does increasing the verification budget rr recover the 0.040.04–0.050.05 GenEval drop (especially in Two-Object and Counting)? What wall-clock speedups remain when matching AR quality, and are these declines consistent across seeds?

  3. Learned spatial drafting vs. 2D refinement. How much of the gain comes from learned spatial initialization versus the row-wise correction schedule itself (e.g., compared with PJD or non-learned initialization under the same verification kernel)?

3 comments

We thank the reviewer for recognizing SSD's ability to substantially accelerate existing image generators and the value of our investigation of latency–quality trade-offs.

Weakness 1: Clarification of the correction rule and reference updates

We are sorry for the confusion. We will clarify the presentation in the revision. SSD selects drafts via argmax rather than probability-based sampling. Draft selection is deterministic given fixed logits.

Reference updates: Across rounds, accepted positions retain their reference distributions, while rejected positions use the current backbone distributions as their next references. We will clarify this mechanism in the revision.

Weakness 2: Quality preservation at the headline speedups

SSD explores acceleration without exact distribution preservation, an important line of research that includes ZipAR [1], GSD [2], LANTERN [3], COOL-SD [4], and MuLo-SD [5]. SSD outperforms ZipAR, the fastest approximate decoding baseline in our paper, while achieving comparable or higher GenEval scores (see the table below). Our approach can be useful when speed is the system's bottleneck and is worth publishing.

ModelLatency (s) ↓\downarrowLatency speedup ↑\uparrowOverall ↑\uparrow
Lumina-mGPT-7B96.231.00×1.00\times0.50
+ ZipAR (w=6w=6)14.556.61×6.61\times0.46
+ SSD (Ours)13.097.35×7.35\times0.46
Janus-Pro-7B7.721.00×1.00\times0.77
+ ZipAR (w=4w=4)1.874.13×4.13\times0.72
+ SSD (Ours)1.455.32×5.32\times0.73
Emu3-8B281.061.00×1.00\times0.45
+ ZipAR (w=4w=4)25.8210.89×10.89\times0.31
+ SSD (Ours)25.0011.24×11.24\times0.40

Question 1: Clarification of the correction rule and reference updates

We thank the reviewer for pointing this out and will clarify these details in the revision. Specifically, SSD selects drafts via argmax rather than sampling from qkq_k.

Definition of qkq_k: qkq_k is the correction reference distribution obtained by dividing the CFG-adjusted draft logits at position kk by the temperature TT and applying softmax over the visual-token vocabulary.

Reference-distribution update rule: Each round re-evaluates the current draft with the backbone; accepted positions retain their reference distributions, while rejected positions replace them with the current backbone distributions.

Question 2: Quality-matched speedups on GenEval

We thank the reviewer for this suggestion. We conducted additional experiments on Janus-Pro-7B and found that increasing the verification budget from r=2r=2 to r=3r=3 improves the GenEval score by 0.01.

With r=3r=3, SSD still achieves a 4.17× wall-clock speedup with a GenEval score closer to that of AR. However, the gap relative to AR's GenEval score persists across seeds. Nevertheless, with the default setting (r=2r=2), SSD achieves a higher GenEval score than ZipAR [1] (0.73 vs. 0.72), while also providing a larger wall-clock speedup (5.32× vs. 4.13×).

MethodVerification rounds rrLatency (s/image) ↓Wall-clock speedup ↑GenEval overall ↑
AR—7.721.00×0.77
ZipAR (w=4w=4)—1.874.13×0.72
SSD (default; paper)21.455.32×0.73
SSD31.8534.17×0.74

Question 3: Learned spatial drafting vs. 2D refinement

We thank the reviewer for this suggestion. Learned initialization improves DPG-Bench by 45.99 points at one verification round and 8.09 points at two rounds compared with random initialization. Its benefit is largest when the correction budget is small.

We compare SSD's learned initialization with random initialization on Janus-Pro-7B, keeping the architecture, backbone, row-wise schedule, and verification kernel unchanged. We report results for r∈{0,1,2}r \in \{0,1,2\}; r=2r=2 is the default setting used in our method.

Verification rounds rrLearned initializationRandom initialization
068.182.73
181.0635.07
282.9774.88

Row-wise correction recovers substantial quality even from random initialization, but does not close the gap within two rounds. We will include this ablation in the revision.

References

[1] He et al. ZipAR: Parallel Autoregressive Image Generation through Spatial Locality. ICML, 2025.

[2] So et al. Grouped Speculative Decoding for Autoregressive Image Generation. ICCV, 2025.

[3] Jang et al. LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding. ICLR, 2025.

[4] Li et al. Annealed Relaxation of Speculative Decoding for Faster Autoregressive Image Generation. AAAI, 2026.

[5] Peruzzo et al. Multi-Scale Local Speculative Decoding for Image Generation. CVPR, 2026.

AuthorsOpenReview ↗

We welcome your feedback.

We thank the reviewer for their time and consideration. We hope our responses have addressed their questions and concerns, and would appreciate any further feedback. We would be happy to clarify any remaining points.

AuthorsOpenReview ↗

Brief Follow-Up

We would like to briefly follow up on our responses, as the rebuttal deadline is approaching. We hope they have addressed your concerns and would appreciate any further feedback or questions you may have.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.