- 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
- 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?
- 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?
- 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?
- 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?
- 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.
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