- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
DyMPQ quantizes diffusion transformers with a shared W4A4 branch and stage-specific W8A8 branches for selected channels. It selects channels by activation RMS, partitions denoising stages by dynamic programming on reconstruction error, and jointly calibrates the branches after a shared channel permutation. Tests cover PixArt-Sigma, Z-Image-Turbo, Z-Image Base, and SANA-Video. On the Z-Image models, it reports 2.21-2.24x single-step speedups and 55.8-61.8% lower peak memory than BF16, while largely retaining generation quality.
Critical strengths
-
Quality holds up across several models. PixArt-Sigma reaches 0.915 ImageReward versus 0.917 for FP16, and Z-Image Base reaches 75.40% GenEval versus 75.53% for BF16 (Figure 4; Table 1). On SANA-Video, DyMPQ leads the reported quantized methods in six of eight VBench dimensions, using 1,255 paired videos per method (Table 2; Appendix B.2).
-
Table 3 directly supports stage adaptation: at the same per-step high-precision channel budget, dynamic selection raises ImageReward from 0.845 to 0.881 and lowers LPIPS from 0.48 to 0.39 before CJC. The fused INT4/INT8 kernels achieve about 2.2x single-step speedup on an RTX 4090, with about 3% single-step latency overhead over naive W4A4 (Figures 6-7).
Critical weaknesses
-
The stage-adaptation ablation is incomplete. Table 3 omits Static MP + CJC, so it does not show whether the dynamic gain remains after both variants receive joint calibration. This comparison is also missing for six-stage Z-Image Base and four-stage SANA-Video. Additional stages store more W8 weights, making storage matching important. Given the precedents in channel-wise precision (Atom), reordering (PermuQuant), and compensation (GPTQ), these controls would clarify the added value. The PixArt ablation supports the mechanism, but not the broader quality/storage claim.
-
SVDQuant has resource results on Z-Image Base but no quality scores (Table 1; Appendix Table 4; Figure 6). It uses 4.0 GB storage and 5.2 GB peak memory, versus 4.5 GB and 6.0 GB for DyMPQ; speedups are 2.09x and 2.24x, respectively. Without its generation scores, the quality/resource trade-off against this close baseline cannot be assessed. The gains over Q-DiT and ConvRot do not settle that comparison.
Questions for authors
-
Why are SVDQuant's Base quality scores missing? Please give GenEval and PRISM-EN for the Figure 6 configuration, or explain the limitation, including any differences in quantized layer scope or calibration budget.
-
How does TCR + Static MP + CJC compare with DyMPQ on PixArt-Sigma and Z-Image Base, at matched per-step channel counts and at similar stored-weight budgets?
-
What calibration prompts, sample counts, per-timestep activation-row budget, and minimum stage length L were used? What are the offline calibration time and peak memory, including interval-cost computation and CJC?
2 comments
Thanks for the rebuttal. It has addressed my concerns. I will keep my score at 3.
Response to Reviewer 8pWN
We thank the reviewer for the constructive feedback. We have added component ablations and expanded the calibration protocol and offline cost analysis.
1. SVDQuant quality results on Z-Image Base.
The public SVDQuant implementation we used does not directly support Z-Image Base. With our best-effort adaptation, we were unable to obtain valid images under W4A4, W4A2, or W3A3. The previously reported Base efficiency results were obtained through performance simulation using the quantized Z-Image-Turbo weights officially released by SVDQuant, solely to assess runtime costs.
2. Static MP + CJC comparisons and weight storage budgets.
Table 3 now includes Static MP + CJC comparisons on PixArt-Σ and Z-Image Base, plus Base GenEval ablations. Both variants use the same per-step high-precision channel budget.
\hline \text{Configuration}&\text{PixArt IR}\uparrow&\text{PixArt LPIPS}\downarrow&\text{Base GenEval}\uparrow\\\\ \hline \text{FP16 / BF16}&0.917&\text{—}&75.53\\\\ \text{Naive W4A4}&-2.010&0.85&48.52\\\\ \text{+ TCR}&-2.264&0.84&68.05\\\\ \text{+ DyMP}&-0.029&0.63&72.39\\\\ \text{+ TCR + Static MP}&0.845&0.48&72.31\\\\ \text{+ TCR + DyMP}&0.881&0.39&74.63\\\\ \text{+ TCR + Static MP + CJC}&0.898&0.37&73.14\\\\ \text{+ TCR + DyMP + CJC}&0.915&0.35&75.40\\\\ \hline \end{array}$$ With CJC applied to both variants, dynamic selection raises PixArt-Σ ImageReward from 0.898 to 0.915 and Base GenEval from 73.14 to 75.40 (+2.26 percentage points). We also compare both methods under matched auxiliary-weight storage on Turbo and Base. Static denotes TCR + Static MP + CJC; Ours denotes full DyMPQ (TCR + DyMP + CJC). Both use the same W4 backbone size and W8 weight budget ($G\times K$), with all other settings unchanged. $$\begin{array}{llrrrr} \hline \text{Model}&\text{Configuration}&G&K&\text{GenEval}\uparrow&\text{Normalized linear latency}\downarrow\\\\ \hline \text{Z-Image-Turbo}&\text{Static}&1&256&76.34&1.09\\\\ \text{Z-Image-Turbo}&\text{Ours}&2&128&76.28&1.00\\\\ \text{Z-Image Base}&\text{Static}&1&768&74.88&1.22\\\\ \text{Z-Image Base}&\text{Ours}&6&128&75.40&1.00\\\\ \hline \end{array}$$ At matched auxiliary-weight storage, quality is comparable on Turbo, while Ours improves Base GenEval by 0.52 percentage points. The expanded static channel set may suffice for Turbo's 8-step trajectory, while Base's 50-step trajectory may benefit more from stage-specific allocation. Ours executes one 128-channel high-precision branch per step, while Static executes all retained high-precision channels. For Base, linear-layer latency is 22% higher with the static channel budget. Notably, the regular auxiliary-weight layout and predetermined switches offer scope to reduce GPU memory overhead. Following LightX2V's blockwise offloading and asynchronous prefetching [1], inactive high-precision branches could remain in host memory. Each block's GPU buffer could then be reused after its final read in the current stage to prefetch next-stage weights while subsequent blocks execute. This could reduce resident GPU memory usage and overlap part of the transfer with computation. **References** [1] LightX2V Contributors. [LightX2V: Light Video Generation Inference Framework](https://github.com/ModelTC/LightX2V). GitHub repository, 2025. **3. Calibration data, sampling, and offline costs.** We have expanded Appendix B.2 with the calibration settings. PixArt-Σ uses the first 64 captions from the COCO training caption list; Z-Image-Turbo and Base share 64 manually written prompts; SANA-Video uses 32 independently written prompts following the descriptive style of VBench. Each prompt uses one random seed to generate a complete denoising trajectory. For interval-cost estimation, we sample 32 evenly spaced activation rows from each calibration sample at every quantized layer and denoising timestep. The minimum stage length is L=2 denoising steps. The calibration samples will be released with the code in our forthcoming public repository. Appendix B.5 reports the complete offline calibration costs, separating CJC from the remaining calibration steps. The latter include statistics collection for TCR and DyMP, interval-cost computation, stage partitioning, and model loading and saving. All evaluated models can be calibrated on a single RTX 4090; we use data parallelism for some models to reduce calibration time. Calibration costs on RTX 4090 GPUs: $$\begin{array}{lrrrrr} \hline \text{Model} & \text{GPUs} & \text{Other (h)} & \text{CJC (h)} & \text{Total (h)} & \text{GB/GPU} \\\\ \hline \text{PixArt-Σ} & 1 & 0.18 & 14.26 & 14.44 & 13.7 \\\\ \text{Z-Image-Turbo} & 2 & 0.10 & 6.46 & 6.55 & 14.9 \\\\ \text{Z-Image Base} & 4 & 0.67 & 12.55 & 13.22 & 17.2 \\\\ \text{SANA-Video} & 4 & 0.75 & 6.25 & 7.00 & 22.2 \\\\ \hline \end{array}$$