acceptodds
Under review as a conference paper at ICLR 2027

GSTAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation

Abstract

Discrete autoregressive models offer a scalable framework for image generation. However, they face a fundamental trade-off: long token sequences preserve representation quality but slow generation, while aggressive spatial compression reduces representation capacity. Based on product quantization, we show that, under a fixed nominal information budget, group size should scale quadratically with the per-axis spatial compression factor. Building on this principle, we propose GSTAR, a spatial–group co-scaling framework that enables 32x and even effective 64x compression while preserving strong reconstruction, understanding, and generation performance. Since larger groups introduce intra-group dependencies that must be modeled efficiently, we analyze their impact on inference latency and generation quality and introduce a semi-autoregressive decoder that combines parallel feature computation with lightweight causal modeling. With these designs, our highly compressed models perform competitively with leading models across five representative text-to-image benchmarks (GenEval, DPG-Bench, PRISM, LongText-Bench, and Qwen-Image-Bench). At 1024x1024 resolution, our model delivers a 20x speedup over prior discrete autoregressive models, generating each image in 2.5 seconds on a single NVIDIA H200. Together, these results show that highly compressed discrete representations can enable high-quality autoregressive generation at low latency.

Primary area:
generative models
Keywords:
autoregressive image generation, visual tokenization, product quantization, efficient inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vBSzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Under a fixed nominal bit budget B = (HW/s²)·G·log₂K, product quantization implies that per-axis compression s and group size G should co-scale as s → αs, G → α²G, cutting the number of global autoregressive steps by α² at unchanged nominal capacity. The authors instantiate this as a 32× tokenizer with G = 16 (GSTAR-Pro-8B) and an "effective 64×" interface built by losslessly packing 2×2 neighbourhoods of quantized positions into one global unit with G = 32 (GSTAR-Flash-1.7B), so no re-quantization occurs. Since large groups make intra-group code dependencies harder, they add a semi-autoregressive depth head: a parallel trunk computes all group slots once and a single-layer causal Transformer conditions each sampled code on previous ones. Evaluation covers five T2I benchmarks, GEdit-Bench for editing, ImageNet ablations on group size and head design, reconstruction PSNR and a multimodal-understanding appendix, all leading to a 2.5 s/image claim on one H200.

Critical strengths

The co-scaling analysis is simple, correct and actually tested. The observation that a fixed budget forces G to grow quadratically in the compression factor reorganises the design space, and the paper does the controlled work rather than asserting it: Table 2 sweeps G at fixed K and fixed 32×, reporting reconstruction, generation and linear-probe accuracy, and Table 3 repeats it at T2I scale. That makes the paper useful even to readers who never adopt the full system.

The absolute system numbers matter for the discrete-AR camp. An 8B discrete model at 50.07 on Qwen-Image-Bench versus 49.89 for Emu3.5 (34.1B) and 49.23 for Qwen-Image (20B), plus GEdit-Bench results beaten only by Emu3.5, is a real result; the understanding appendix (64.7 vs 64.6 against continuous Qwen-ViT features) defuses the usual objection that quantized semantic tokens destroy comprehension.

Critical weaknesses

The paper's central claim is contradicted by its own numbers. The abstract and Section 4.2 say co-scaling preserves quality while shortening the sequence, but Table 2 shows gFID 5.74 → 17.05 going from G=16 to G=32, and Appendix D shows the packed 64× G=32 model is worse than its unpacked 32× G=8 source on all three metrics (0.7885 / 80.44 / 5.13 vs 0.8060 / 82.49 / 5.30) despite being, by construction, the same codes under a different global ordering. Table 3 likewise puts G=16 barely above G=8 (PRISM 5.46 vs 5.30). The real finding is that extra per-position capacity helps reconstruction and semantics but the generator cannot exploit it, and that packing four positions into one step costs quality. That is a defensible finding — but it is not the one advertised, and the "effective 64×" framing overstates what is supported, which matters because the 32×→64× design story rests on it.

The latency claim is not decomposable and most of the win is engineering. Figure 5 attributes 8.9× end-to-end to four cumulative steps, of which CUDA-graph capture, Triton kernel fusion and SGLang serving are engineering; only the semi-autoregressive head (1.49×) is a scientific contribution. The abstract separately claims 20× over prior discrete AR models and 2.5 s/image against an unstated baseline, with no statement of whether VAE decode and the DiT refiner (Appendix A.2, initialised from Ideogram 4) are included. Since that refiner is a diffusion model conditioned on the discrete tokens, a reader cannot tell how much of the reported fidelity comes from the tokenizer/generator and how much from the diffusion decoder — which also muddies the comparison against diffusion baselines in Table 1.

Questions for authors
  1. Reconcile the 20× in the abstract with the 8.9× in Section 5.3: measured against what, on what hardware, and does 2.5 s include the DiT refiner and VAE decode?
  2. Table 8 shows the packed 64× model worse than its 32× source on all three metrics, although packing is lossless and only reorders global steps. What explains this, and does it not undercut the Section 4.2 claim?
  3. In Table 2 gFID degrades sharply from G=16 to G=32 while rFID improves. How was this resolved for GSTAR-Flash (G=32), and did the 64× model need more steps to converge, as Figure 4 suggests?
  4. Table 1 compares an 8B model against 14B–34B systems. What were the training data and compute, and what does the comparison look like at matched compute rather than matched parameter count?
  5. Does the semi-autoregressive head's 1.49× advantage survive once CUDA-graph capture is also applied to the causal head? The measured gain predates the graph and fusion optimisations, which change the relative cost of the sequential loop substantially.
1 comment

Dear reviewer vBSz,

We greatly appreciate your time and effort in reviewing our work. We are eager to ensure that we have adequately addressed your concerns and are prepared to offer further clarifications or address any additional questions you may have. We would be grateful if you could share your thoughts on our rebuttal.

Best regards, The Authors

Reviewer_pku5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GSTAR shortens the global spatial sequence while increasing the number of discrete codes per position. A native 32x grouped tokenizer is converted to an effective 64x interface through lossless 2x2 packing. A semi-autoregressive head combines parallel feature computation with a lightweight causal layer to model dependencies within each code group. Experiments cover reconstruction, understanding, text-to-image generation, and editing, with a visual refiner used in the generation system. The Flash model reports 2.5-second generation at 1024 x 1024 on one H200.

Critical strengths
  1. The group decoder addresses a relevant bottleneck. In Table 4, the Semi-AR head matches the causal head's PRISM score of 5.46 while reducing aggregate head latency at 512 x 512 from 6.44 to 3.41 seconds.
  2. The paper evaluates reconstruction, generation, editing, and understanding, and distinguishes nominal bit capacity from empirical information and invertible packing from additional lossy quantization.
Critical weaknesses
  1. When additional group capacity helps generation: In Table 2's ImageNet setting, gFID is 2.1057 for G=1 and 5.7388 for G=16 despite improved reconstruction. The paper acknowledges slower convergence; Figure 4 compares only G=8/16, while Table 3 supports G=16 in text-to-image generation. What interpretation of the G=1 versus G=16 difference is supported by the reported curves and any available training logs?
  2. Interpretation of the head comparison: Figure 3 describes the parallel trunk and lightweight causal layer, but Table 4 does not give the numeric configurations or parameter budgets of the compared heads. What widths, depths, and parameter counts were used for each head in this table? These details are needed to interpret the head-latency comparison.
  3. Attribution of the acceleration: Figure 1(d) reports roughly 21x Flash speed ratios, while Figure 5 gives a separate 8.9x internal optimization sequence. Appendix C states matched timing conditions, but the baseline implementation sources and 1024 x 1024 component times are unclear; Table 8 reports packing quality without paired latency. How do the existing measurements support each claimed source of acceleration? Please identify the baseline implementations and timing sources, and use any available component logs to distinguish measured gains from the packing benefits expected from the shorter global sequence.
  4. Evaluation of the decoding stack: Appendix A.2's Ideogram 4-initialized refiner conditions only on visual codes, without extra text. Figure 6 compares it qualitatively with the ViT decoder, but Table 9 does not identify which produces its PSNR values. Which decoder is used in each quantitative evaluation, and what do the existing same-code examples show about the refiner's contribution?
Questions for authors

Please refer to weaknesses.

1 comment

Dear reviewer pku5,

We greatly appreciate your time and effort in reviewing our work. We are eager to ensure that we have adequately addressed your concerns and are prepared to offer further clarifications or address any additional questions you may have. We would be grateful if you could share your thoughts on our rebuttal.

Best regards, The Authors

Reviewer_PQ68OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GSTAR, which shortens the global autoregressive generation sequence by reducing the number of spatial positions in the image while increasing the number of discrete codes assigned to each position. It further introduces a semi-autoregressive head that combines parallel feature computation with lightweight causal prediction, thereby reducing the cost of generating multiple codes within each position. Experiments show that the method achieves competitive quality on several text-to-image benchmarks and can generate a single 1024×1024 image in as little as 2.5 seconds on one H200 GPU.

Critical strengths

The paper presents a detailed and systematic optimization process, including CUDA Graphs, operator fusion, and SGLang. It demonstrates the cumulative speedup achieved by these optimizations rather than reporting only theoretical computational complexity. The overall engineering and systems work is thorough and provides a complete account of the acceleration pipeline.

Critical weaknesses

Insufficient discussion of the difference from RQ-Transformer. The paper does not clearly explain how GSTAR differs from RQ-Transformer. Both approaches first use a relatively large model to generate spatial tokens and then generate local tokens. If the main contribution is essentially parallel feature computation combined with a small autoregressive Transformer, the novelty of the work appears limited.

Incomplete speed comparisons under different compression ratios. The comparison in Table 1 is unusual: it evaluates performance under different compression ratios, but does not report the corresponding speedups at those compression ratios. This makes the experimental analysis incomplete. In Figure 1(d), the reported speed comparison appears to use a model without compression as the baseline, rather than comparing against models with the same compression ratios.

Insufficient analysis of the effect of gg. Table 2 reveals an interesting phenomenon: increasing gg makes the generation task more difficult to learn, while making the reconstruction task easier. However, the paper does not further explain how it alleviates the negative impact of gg on generation learning.

Questions for authors

If a larger gg significantly degrades the model’s ability to learn image generation, why does the final generation quality in Table 1 remain comparable to that of existing compressed models?

Can the proposed strategy achieve higher generation speed than existing open-source compressed models? The paper does not appear to report speed comparisons against compressed models. If GSTAR is not faster than existing compressed models, what is the practical significance of the proposed method?

1 comment

Dear reviewer PQ68,

We greatly appreciate your time and effort in reviewing our work. We are eager to ensure that we have adequately addressed your concerns and are prepared to offer further clarifications or address any additional questions you may have. We would be grateful if you could share your thoughts on our rebuttal.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.