acceptodds
Under review as a conference paper at ICLR 2027

DyMPQ: Dynamic Mixed-Precision Quantization for Efficient Diffusion Transformers

Abstract

Diffusion transformers (DiTs) have demonstrated strong performance in image and video generation, but their large model sizes and iterative denoising incur substantial computational and memory costs. However, timestep-dependent activation outliers and channel-wise magnitude disparities hinder accurate 4-bit quantization. To address these challenges, we propose Dynamic Channel-wise Mixed-Precision Quantization (DyMPQ) for DiTs. It combines a shared W4A4 main branch with stage-specific high-precision branches to absorb timestep-varying activation outliers. We then apply trajectory-based channel reordering to reduce within-group magnitude disparities in the W4A4 main branch. Building on this mixed-precision representation, we further introduce cross-stage joint calibration to reduce weight quantization error and kernel fusion to enable efficient inference. Extensive experiments on PixArt-, Z-Image, and SANA-Video demonstrate that our method effectively preserves image and video generation quality under low-bit quantization. On an NVIDIA RTX 4090, our method achieves up to single-step speedup and reduces peak GPU memory by up to on Z-Image compared with BF16.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Diffusion Transformers, Post-Training Quantization, Mixed-Precision 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_8pWNOpenReview ↗
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
  1. 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).

  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
  1. 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.

  2. 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
  1. 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.

  2. 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?

  3. 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
AuthorsOpenReview ↗

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}$$

Thanks for the rebuttal. It has addressed my concerns. I will keep my score at 3.

Reviewer_unkXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DyMPQ is a post-training quantization method for diffusion transformers. Its shared W4A4 with stage specific W8A8 branches handle the changing outliers. They use channel reordering to improve low-bit grouping and joint calibration to compensate quantization errors of weights. Authors also implement fused kernels for PixArt-Sigma, Z-Image and SANA-Video and evaluation metrics. Finally, they accelerate Z-Image more than 2x on RTX 4090 over BF16 precision.

Critical strengths
  1. The proposed representation addresses temporal variation while retaining a shared low-bit backbone. Stage-specific FP8 channels provide flexibility without storing a separate full model for each stage. Table 3 supports this design: with channel reordering and the same per-step high-precision channel budget, dynamic selection improves LPIPS and ImageReward scores over static selection.
  2. The paper connects quantization quality to an executable implementation. The fused implementation and reported inference measurements strengthen the practical contribution. This matters because additional branches and activation preparation can otherwise negate the benefits of low-bit computation.
Critical weaknesses
  1. OWQ already adjusts retained high-precision columns to compensate for errors introduced when quantizing other columns. Thus, allowing the W8 branch to absorb W4 weight-quantization error is not itself a new principle. The more specific contribution is the joint treatment of shared W4 weights and stage-specific W8 branches. However, Table 3 does not establish whether the dynamic advantage persists when the static configuration also receives corresponding error-compensation calibration.
  2. Channels are selected from aggregated calibration statistics and remain fixed within each stage across prompts and seeds. Temporal variation alone does not establish that these selections adequately cover important channels on unseen inputs. Separate calibration and evaluation prompts provide useful indirect evidence, but the paper lacks a direct analysis of selection robustness.
  3. Increasing the number of stages increases W8 storage even though each step executes only one branch. Without a stage-count sensitivity analysis, it remains unclear how much benefit requires additional stored weights and whether the chosen configurations offer a favorable trade-off.
Questions for authors
  1. Does dynamic channel selection retain its advantage when both static and dynamic configurations use corresponding GPTQ-based branch compensation, with the same per-step high-precision channel budget?
  2. On held-out prompts and seeds, how closely does the offline channel selection approach a per-example Top-K selection in terms of activation reconstruction error?
  3. How do generation quality, total weight storage, inference latency, and calibration cost change with the number of stages? In particular, what evidence motivated the reported stage counts, and does finer partitioning provide meaningful gains over the chosen configurations?
2 comments
AuthorsOpenReview ↗

Response to Reviewer unkX

We thank the reviewer for the constructive feedback. We have added static baselines with error compensation, an analysis of offline channel selection generalization, and stage-count experiments to address the concerns below.

1. Benefits of dynamic selection when the static configuration also receives error compensation.

We have added Static MP + CJC baselines for PixArt-Σ and Z-Image Base to Table 3. We apply GPTQ-based error compensation to both static and dynamic configurations under the same per-step high-precision channel budget.

\hline \text{Configuration}&\text{Zimage-Base GenEval}\uparrow&\text{PixArt IR}\uparrow&\text{PixArt LPIPS}\downarrow\\\\ \hline \text{FP16 / BF16}&75.53&0.917&\text{--}\\\\ \text{Naive W4A4}&48.52&-2.010&0.85\\\\ \text{+ TCR}&68.05&-2.264&0.84\\\\ \text{+ DyMP}&72.39&-0.029&0.63\\\\ \text{+ TCR + Static MP}&72.31&0.845&0.48\\\\ \text{+ TCR + DyMP}&74.63&0.881&0.39\\\\ \text{+ TCR + Static MP + CJC}&73.14&0.898&0.37\\\\ \text{+ TCR + DyMP + CJC}&75.40&0.915&0.35\\\\ \hline \end{array}$$ With CJC calibration, the dynamic configuration still improves Base GenEval by 2.26 percentage points over the static configuration and increases PixArt ImageReward from 0.898 to 0.915. These results show that the benefits of dynamic selection persist when the static configuration also uses error compensation. **2. Generalization of offline channel selection.** In Appendix C.3, we compare offline and per-sample channel selection on 32 GenEval prompts not used for calibration, with one random seed per prompt. Offline selection reuses the Top-128 channels determined for each layer and stage during calibration; per-sample selection reselects the Top-128 channels at each stage for the current sample. The activation reconstruction errors are reported below (NMSE in units of $10^{-3}$; lower is better). PixArt-Σ $$\begin{array}{lcc} \hline \text{Channel selection}&\text{Stage 1}&\text{Stage 2}\\\\ \hline \text{Pure A4}&26.740&26.914\\\\ \text{Offline Top-128 (ours)}&6.297&7.740\\\\ \text{Per-sample Top-128}&5.458&6.738\\\\ \hline \end{array}$$ Z-Image Base $$\begin{array}{lcccccc} \hline \text{Channel selection}&\text{Stage 1}&\text{Stage 2}&\text{Stage 3}&\text{Stage 4}&\text{Stage 5}&\text{Stage 6}\\\\ \hline \text{Pure A4}&17.366&17.207&17.030&16.976&16.966&17.171\\\\ \text{Offline Top-128 (ours)}&4.381&4.488&4.613&5.005&5.834&6.486\\\\ \text{Per-sample Top-128}&4.256&4.347&4.452&4.831&5.565&6.217\\\\ \hline \end{array}$$ The offline approach substantially reduces reconstruction error relative to pure A4 at every stage and approaches the results of per-sample reselection, supporting its generalization to held-out inputs. **3. Trade-offs among stage count, storage, quality, and cost.** In Appendix C.4 and Figure 11, we compare different stage counts on Z-Image Base with a fixed high-precision channel budget. The table reports generation quality, single-step latency, and estimated calibration time. $$\begin{array}{cccc} \hline G&\text{GenEval}\uparrow&\text{Latency (ms)}&\text{Calibration (h)}\\\\ \hline 1&73.14&493.74&13.22\\\\ 2&74.09&493.44&13.23\\\\ 4&74.53&494.84&13.25\\\\ 6&75.40&498.94&13.22\\\\ 8&75.10&496.84&13.19\\\\ 10&75.50&499.64&13.23\\\\ 12&75.40&500.22&13.27\\\\ \hline \end{array}$$ Increasing $G$ from 1 to 6 improves GenEval by 2.26 percentage points, while finer partitions provide at most an additional 0.10 percentage points. Although stage-specific W8 weight storage grows linearly with $G$ and runtime GPU memory usage increases (Figure 11(b)), single-step latency remains approximately constant because each step executes only its current stage's high-precision branch. Calibration cost also remains stable, as the calibration data and amount of trajectory replay are unchanged. These results support our choice of $G=6$.

Thank you for the targeted additional experiments. The compensated static baseline addresses my main concern about attributing the gains to dynamic selection. The held-out reconstruction analysis and stage-count study also substantially strengthen the evidence. The robustness evaluation remains limited to 32 prompts with one seed each, and the revision should clearly quantify the additional weight storage separately from runtime GPU memory. Overall, these additions sufficiently alleviate my main concerns, and I am inclined to raise my rating to 3.

Reviewer_bC7kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DyMPQ, a post-training quantization framework for diffusion transformers. It combines stage-specific high-precision channels, trajectory-based channel reordering, and cross-stage joint calibration to address timestep-varying activation distributions. Fused kernels support efficient mixed-precision execution. Experiments cover several image and video models, with generation-quality evaluation and single-step efficiency measurements

Critical strengths
  1. The paper addresses the practical challenge of timestep-varying activation distributions with a coherent mixed-precision design.
  2. The implementation includes real fused kernels and evaluations across multiple models, demonstrating substantial engineering effort.
Critical weaknesses
  1. The advantage over strong baselines is not consistently clear. DyMPQ preserves generation quality well, but some gains over strong baselines are small and are reported without uncertainty estimates. SVDQuant quality results are also missing on Z-Image Base, although its efficiency results are reported.

  2. The existing semantic evaluation is not sufficiently connected to the method’s components. The paper reports semantic-related results and qualitative examples, but the component ablations do not show how stage-specific channel selection, channel reordering, and joint calibration contribute to semantic preservation.

  3. The dynamic design is not compared at matched total storage. The static and dynamic variants use the same per-step high-precision channel budget, but the dynamic method stores multiple stage-specific branches. Its advantage at an equal total storage budget therefore remains unclear.

  4. Deployment costs are not fully characterized. The paper reports single-step inference efficiency, while calibration cost and complete-generation latency are not reported. These details would help assess the practical trade-off.

Questions for authors
  1. The advantage over strong baselines is not consistently clear. DyMPQ preserves generation quality well, but some gains over strong baselines are small and are reported without uncertainty estimates. SVDQuant quality results are also missing on Z-Image Base, although its efficiency results are reported.

  2. The existing semantic evaluation is not sufficiently connected to the method’s components. The paper reports semantic-related results and qualitative examples, but the component ablations do not show how stage-specific channel selection, channel reordering, and joint calibration contribute to semantic preservation.

  3. The dynamic design is not compared at matched total storage. The static and dynamic variants use the same per-step high-precision channel budget, but the dynamic method stores multiple stage-specific branches. Its advantage at an equal total storage budget therefore remains unclear.

  4. Deployment costs are not fully characterized. The paper reports single-step inference efficiency, while calibration cost and complete-generation latency are not reported. These details would help assess the practical trade-off.

  5. As an optional extension, testing a 2-bit main branch while retaining the high-precision auxiliary branches could further examine semantic preservation under more aggressive quantization.

2 comments
AuthorsOpenReview ↗

Response to Reviewer bC7k

We thank the reviewer for the constructive feedback.

1. Advantages over strong baselines and SVDQuant results on Z-Image Base.

At W4A4, DyMPQ's gains are modest on Z-Image-Turbo but larger on Base, where its GenEval score is 75.40 versus 75.53 for BF16. Gains increase at W3A3 and W4A2.

\hline \text{Model}&\text{Main-branch precision}&\text{Baseline}&\text{Baseline GenEval}\uparrow&\text{DyMPQ}\uparrow&\text{Gain (percentage points)}\\\\ \hline \text{Z-Image-Turbo}&\text{W4A4}&\text{SVDQuant}&75.94&76.28&+0.34\\\\ \text{Z-Image-Turbo}&\text{W3A3}&\text{ConvRot}&70.99&76.28&+5.29\\\\ \text{Z-Image-Turbo}&\text{W4A2}&\text{ConvRot}&62.61&71.05&+8.44\\\\ \text{Z-Image Base}&\text{W4A4}&\text{ConvRot}&70.13&75.40&+5.27\\\\ \text{Z-Image Base}&\text{W3A3}&\text{ConvRot}&16.93&55.77&+38.84\\\\ \text{Z-Image Base}&\text{W4A2}&\text{ConvRot}&4.36&21.37&+17.01\\\\ \hline \end{array}$$ Detailed quantization settings are provided in Appendix B.1, with complete results in Table 1 of the revised paper. The public SVDQuant implementation we used does not directly support Z-Image Base. With our best-effort adaptation, we were unable to obtain valid images under W4A4, W4A2, or W3A3. The previously reported SVDQuant efficiency results for Base were obtained through performance simulation using the quantized Z-Image-Turbo weights officially released by SVDQuant, solely to assess runtime costs. **2. Contributions of individual components to semantic preservation.** Table 3 includes Base GenEval ablations and Static MP + CJC comparisons for PixArt-Σ and Base. TCR and stage-specific selection provide complementary benefits; CJC further improves quality, and dynamic selection retains its advantage when both configurations use CJC. $$\begin{array}{lrrr} \hline \text{Configuration}&\text{Base GenEval}\uparrow&\text{PixArt ImageReward}\uparrow&\text{PixArt LPIPS}\downarrow\\\\ \hline \text{FP16 / BF16}&75.53&0.917&\text{—}\\\\ \text{Naive W4A4}&48.52&-2.010&0.85\\\\ \text{+ TCR}&68.05&-2.264&0.84\\\\ \text{+ DyMP}&72.39&-0.029&0.63\\\\ \text{+ TCR + Static MP}&72.31&0.845&0.48\\\\ \text{+ TCR + DyMP}&74.63&0.881&0.39\\\\ \text{+ TCR + Static MP + CJC}&73.14&0.898&0.37\\\\ \text{+ TCR + DyMP + CJC}&75.40&0.915&0.35\\\\ \hline \end{array}$$ **4. Calibration cost and end-to-end generation latency.** Appendix B.5 separates CJC from other calibration steps, including TCR/DyMP statistics collection, interval-cost computation, stage partitioning, and model loading/saving. All evaluated models can be calibrated on one RTX 4090; data parallelism reduces calibration time for some models. $$\begin{array}{lrrrrr} \hline \text{Model}&\text{RTX 4090 GPUs}&\text{Other calibration (h)}&\text{CJC (h)}&\text{Total (h)}&\text{GPU memory (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}$$ Appendix C.2 reports end-to-end latency at 1024×1024 on one RTX 4090. DiT forward passes dominate generation latency, with linear layers accounting for a major share of the BF16 forward-pass cost. Quantization and kernel fusion therefore reduce full-generation latency. Timing includes text encoding, all denoising steps, sampler updates, VAE decoding, and output processing, excluding model loading and file I/O. $$\begin{array}{lrrr} \hline \text{Model}&\text{BF16}&\text{DyMPQ}&\text{Speedup}\\\\ \hline \text{Z-Image-Turbo, 8 steps}&\text{4.088 s}&\text{1.938 s}&\text{2.11×}\\\\ \text{Z-Image Base, 50 steps}&\text{51.913 s}&\text{23.391 s}&\text{2.22×}\\\\ \hline \end{array}$$ **5. Semantic preservation under more aggressive quantization.** We have added W3A3 and W4A2 evaluations for Z-Image and W3A3 evaluations for SANA-Video. Detailed quantization settings are provided in Appendix B.1, with image and video results in Tables 1 and 2, respectively. Qualitative comparisons for Z-Image-Turbo at lower bit widths are provided in Figures 6, 15, and 16. $$\begin{array}{llrrrr} \hline \text{Model}&\text{Main-branch precision}&\text{BF16 reference}&\text{ConvRot}&\text{DyMPQ}&\text{Gain over ConvRot (percentage points)}\\\\ \hline \text{Z-Image-Turbo}&\text{W3A3}&75.37&70.99&76.28&+5.29\\\\ \text{Z-Image-Turbo}&\text{W4A2}&75.37&62.61&71.05&+8.44\\\\ \text{Z-Image Base}&\text{W3A3}&75.53&16.93&55.77&+38.84\\\\ \text{Z-Image Base}&\text{W4A2}&75.53&4.36&21.37&+17.01\\\\ \hline \end{array}$$ On SANA-Video at W3A3, DyMPQ achieves an Overall Consistency score of 27.02, close to BF16's 27.18 and above SVDQuant's 26.60 and ConvRot's 21.56. Its Imaging Quality score is 65.19, compared with 63.14 for SVDQuant and 57.62 for ConvRot. These results support DyMPQ's quality advantage over the evaluated baselines at lower bit widths.
AuthorsOpenReview ↗

Response to Reviewer bC7k-Q3

3. Comparison under matched weight storage budgets.

We added comparisons on Z-Image-Turbo and Z-Image Base, keeping the shared W4 backbone size and total W8 storage (G×KG\times K) identical between static and dynamic configurations. All other settings remain unchanged.

\hline \text{Model}&\text{Configuration}&G&K&\text{GenEval}\uparrow&\text{Normalized linear latency}\downarrow\\\\ \hline \text{Z-Image-Turbo}&\text{Static MP}&1&256&76.34&1.09\\\\ \text{Z-Image-Turbo}&\text{DyMP}&2&128&76.28&1.00\\\\ \text{Z-Image Base}&\text{Static MP}&1&768&74.88&1.22\\\\ \text{Z-Image Base}&\text{DyMP}&6&128&75.40&1.00\\\\ \hline \end{array}$$ At matched auxiliary-weight storage, quality is comparable on Turbo, while DyMP improves Base GenEval by 0.52 percentage points. This difference may arise because the expanded static channel set suffices for Turbo's 8-step trajectory, whereas stage-specific channel allocation helps adapt to activation distributions at different stages of Base's 50-step trajectory. Meanwhile, DyMP executes only one auxiliary branch with 128 high-precision channels per step, while the static configuration 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 branch switches offer scope to reduce the dynamic design's additional GPU memory overhead. LightX2V's blockwise offloading and asynchronous prefetching mechanisms [1] could be adapted to keep inactive high-precision branches 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.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Discussion

Dear Area Chair,

We thank you and the reviewers for your time and constructive feedback. DyMPQ introduces a stage-adaptive channel-wise mixed-precision framework for diffusion transformers, integrating channel reordering and cross-stage joint calibration to preserve generation quality under low-bit quantization, with fused kernels enabling efficient inference.

During the discussion, Reviewer unkX stated that the compensated static baseline and additional analyses sufficiently alleviated their main concerns and raised their rating from 2 to 3. Reviewer 8pWN confirmed that the rebuttal addressed their concerns and maintained their rating of 3. The additional experiments strengthen the evidence in three main aspects:

  1. DyMPQ’s quality advantages become more pronounced at lower bit widths. The additional W3A3 and W4A2 evaluations show larger gains over strong baselines. On Z-Image-Turbo, DyMPQ retains near-BF16 generation quality at W3A3 and outperforms ConvRot by 8.44 percentage points in GenEval at W4A2. Additional results on Z-Image Base and SANA-Video also demonstrate quality advantages over quantized baselines, further supporting the method’s effectiveness across image and video generation.

  2. Stage adaptation provides additional quality benefits. We added Static MP + CJC comparisons on PixArt-Σ and Z-Image Base. With the same per-step high-precision channel budget and both variants using channel reordering and corresponding calibration, dynamic selection improves Base GenEval from 73.14 to 75.40, further establishing the benefit of stage-specific channel allocation.

  3. Dynamic allocation retains a favorable quality–efficiency trade-off under matched auxiliary-weight storage. With the same shared W4 backbone size and W8 auxiliary-weight budget, the two variants achieve comparable quality on Turbo. On Base, the dynamic variant achieves a higher GenEval score than the static variant (75.40 vs. 74.88), while the static variant incurs 22% higher linear-layer latency. In addition, DyMPQ’s predetermined stage switches offer scope for future on-demand loading and asynchronous prefetching, which could further reduce resident GPU memory usage.

We also added a held-out analysis of offline channel selection, an evaluation of how stage count affects quality and resource usage, and measurements of calibration costs and full-generation latency, providing a more complete characterization of robustness and practical costs.

We have provided point-by-point responses to Reviewer bC7k and have not yet received further feedback. We would appreciate your consideration of the additional evidence and discussion when assessing the paper’s contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.