- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies post-training quantization of coarse-to-fine autoregressive visual generation (ARVG) models. The authors identify two forms of heterogeneity that uniform PTQ ignores: rare large-magnitude tail weights matter disproportionately for quality but are under-weighted by a plain reconstruction loss, and early generation stages are far more sensitive to activation quantization than late stages, even though late stages dominate the FLOPs. They propose HGA-Quant, a training-free framework with two components. T-BGR extends the BGR weight quantizer of DVD-Quant by up-weighting coefficients beyond in the per-channel reconstruction loss when searching the scale and zero point. CAPS assigns A8 to the longest prefix of early stages that fits a FLOPs-weighted bit budget and A4 to the remaining stages. Experiments on Infinity-2B, HART, VAR-d30, and FlexVAR show that T-BGR avoids collapse at W4A4, where most baselines fail. On Infinity-2B, CAPS at an effective W4A5.6 slightly outperforms uniform W4A6, but it lowers quality on the other three backbones.
Critical strengths
-
The method is simple, cheap, and deployment-friendly, and it substantially improves robustness in the most aggressive setting. T-BGR only changes the weighting of the reconstruction objective inside an existing grid search, so the quantized model keeps a standard per-channel INT4 weight representation without low-rank or full-precision outlier branches. Despite this simplicity, it keeps all four backbones usable at W4A4, where RTN, SmoothQuant, SVDQuant, and PTQ4ARVG collapse (e.g., VAR-d30 FID 5.98 vs. 16.94 for QuaRot). This is the main source of practical value in the paper, since W4A4 is an important regime for edge deployment.
-
The coarse-to-fine precision prior is well motivated and supported by a carefully controlled experiment. The budget-matched placement study (Table 4, Figure 4) compares early-high and late-high schedules with nearly identical high-precision FLOPs, yet ImageReward differs by about 0.39. Random schedules with a higher average bit-width still trail the early-high schedule. This cleanly separates where precision is spent from how much is spent. The finding is likely to be useful for future quantization and caching work on next-scale prediction models, beyond this specific method.
Critical weaknesses
-
It is difficult to attribute the reported gains to the proposed tail-aware objective. In the ablation (Table 3), the pipeline without BGR already reaches ImageReward 0.821 at W4A6, compared with 0.770 for QuaRot in Table 1, and T-BGR with adds only about 0.008 on top. Plain BGR () is worse than no BGR. The curve is non-monotonic, and W4A4 prefers . The W8A8 rows (five of six quantized methods beat FP on GenEval) and Figure 4 (CAPS beats uniform W4A8) suggest a noise level comparable to several claimed improvements. The comparison set also omits the most relevant baselines: GPTQ, DVD-Quant from which BGR is taken, and ARVG-specific PTQ methods such as Shift-and-Sum and LiteVAR. Without these, it is unclear whether the improvement comes from T-BGR or from differences in the base pipeline, such as quantizer granularity, Hadamard setup, or the weight quantizer.
-
The efficiency and generality claims for CAPS are stronger than what the evidence establishes. The reported "W4A5.6" is a FLOPs-weighted proxy over linear layers. It is compared against uniform A6 baselines, which are not natively supported by common hardware, and no latency, memory, or precision-switching overhead is measured. CAPS also lowers quality on three of the four backbones. On FlexVAR, for example, FID rises from 2.92 to 8.61, worse than QuaRot's 4.17. Even on Infinity-2B, the result (ImageReward 0.835 vs. 0.955 for FP) does not support the abstract's "near full-precision" claim. Step-wise bit allocation has close precedents in diffusion quantization, such as DVD-Quant's timestep bit switching and TMPDiff, which the paper does not discuss. As a result, the significance of CAPS as a contribution is hard to judge.
Questions for authors
-
What exactly differs between the "no BGR" configuration in Table 3 and the QuaRot baseline, in terms of weight quantizer (RTN vs. GPTQ), granularity, quantized operators, and Hadamard placement? Could the authors report no-BGR, , and on all four backbones with mean ± std over several seeds? This would clarify how much of the improvement is attributable to T-BGR itself.
-
How was selected, and was it tuned on the same Infinity ImageReward evaluation reported in the main table? This would clarify whether the main results reflect test-set tuning.
-
What are the measured end-to-end latency and peak memory of W4A8, W4A4, and CAPS with real integer kernels, including the cost of switching activation precision at stage boundaries? This would clarify whether the effective-bit reduction translates into practical speedup.
The authors have not yet submitted a rebuttal near the end of the discussion period. My concerns about how much the gains attributable to the proposed components is unclear and claims exceeding the experimental evidence remain unresolved. I therefore maintain my recommendation of weak rejection.