Pith. sign in

REVIEW 4 major objections 5 minor 46 references

TASQ: Temporal-Adaptive Bit Sparsification Quantization for Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read TASQ stores one shared high-precision weight buffer and learns a Temporal-Spatial LSB Mask, so diffusion models execute each denoising stage at its own effective bit-width and cut BitOPs by 25–50% over static quantization without quality…

desk verdict TASQ is a genuinely new temporal weight-precision method for diffusion models, with solid empirical support and one load-bearing representational assumption that deserves scrutiny before the cycle-savings claims are taken at face value. read the letter →

arxiv 2608.03057 v1 pith:247LU6SX submitted 2026-08-04 cs.CV

Seokho Han , Dongwei Wang , Jinhee Kim , Yiran Chen , Kang Eun Jeon , Huanrui Yang , Jong Hwan Ko This is my paper · ORCID
classification cs.CV
keywords temporal-adaptivequantizationdiffusionmodelsbitsparsificationLSBmaskbit-serialexecutionmixedprecisionquantization-awaretrainingdenoisingstages
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Static quantization of diffusion models forces every denoising step to pay the bit cost of the most sensitive step, wasting computation on stages that tolerate lower precision. TASQ separates storage from operating precision: it keeps one shared high-precision weight buffer and learns a Temporal-Spatial LSB Mask that truncates least-significant bits per layer and per denoising stage, so fewer bits are used exactly where the noise schedule permits. A co-designed bit-serial engine executes only the selected bit planes, with no per-stage weight copies or runtime search and no measured cycle overhead when precision changes. On PixArt-Sigma, SANA-1.6B, and SDXL-Turbo, TASQ keeps image quality comparable to static W8/W4 quantization while reducing BitOPs, and together with the engine it cuts execution cycles by 25–50% over static quantization.

What carries the argument

The central object is the Temporal-Spatial LSB Mask $M_{t,l} \in [0,1]$, a per-layer, per-stage gate that multiplies the extracted least-significant bit of a shared quantized weight. The operating weight is $Q^{(t,l)}_{M,n} = Q_n - (1-M_{t,l})\cdot\text{LSB}$, so $M=1$ keeps the bit and $M=0$ drops it; recursively applied, the mask selects how many bit planes each layer reads at each stage. Accompanying it are the Farthest-Stage-First update order, which keeps the shared MSBs from overfitting to neighboring stages, and the Temporal-Precision Engine, an activation-stationary bit-serial dataflow based on the BISMO scheduler that streams only the selected weight planes and therefore has execution cost proportional to effective precision.

What would settle it

Take a diffusion model and compare TASQ against per-stage independent re-quantization at matched average bit-width: for each layer and stage, quantize the full-precision weights to the same effective precision without requiring the stage weights to be truncations of a single buffer. If the per-stage re-quantized model achieves meaningfully better FID/ImageReward than TASQ at the same mean BitOPs, the shared-buffer truncation constraint is the limiting factor and the paper's central premise is testably false; if TASQ matches it, the prefix constraint is harmless in practice.

Watch

Extended reading notes

Core claim

The paper claims that one shared maximum-precision weight buffer is enough to realize near-optimal per-stage operating precisions. From a quantized weight $Q_n$, the $(n-1)$-bit weight is obtained by a right shift (Eq. 1), so dropping the extracted LSB $Q_n - 2Q_{n-1}$ under a learned mask $M_{t,l}$ (Eq. 4) yields stage- and layer-specific effective bit-widths without re-quantization or extra storage. Trained with LoRA-based distillation, a noise-matching loss, and an $\ell^1$ penalty on retained LSBs, the masks anneal to binary decisions; a farthest-stage-first update order prevents the shared MSBs from overfitting to adjacent stages. On DiT and U-Net backbones, this yields quality comparable to or better than static quantization at the same average bit budget, with the learned allocation tracking a closed-form timestep-sensitivity score derived from DDIM error propagation.

Load-bearing premise

Every stage's operating weights are assumed to be well-approximated by truncating the least-significant bits of one common high-precision quantized weight; if the best low-precision weight for a late stage has a different most-significant-bit pattern than the best weight for an early noisy stage, the shared-prefix constraint prevents TASQ from reaching the quality–compute tradeoff of per-stage re-quantization.

Editorial extensions

If this is right

  • Storage precision and operating precision are decoupled: one worst-case buffer serves every stage, so memory cost is set once while arithmetic cost is paid per stage.
  • Execution cycles on precision-scalable hardware follow the average effective bit-width, giving 25–50% cycle reductions over static quantization at equal quality.
  • Spatial-only and temporal-only adaptivity are complementary; joint masks outperform either alone (PixArt-Sigma W4A4 FID 16.1 vs 17.1 and 18.4).
  • The scheme transfers across DiT and U-Net architectures, including a 4-step SDXL-Turbo, so temporal allocation helps even with few denoising steps.
  • Training overhead is small: with the same 5k-iteration schedule, TASQ matches QAT in wall-clock time and memory, and a further 1.4k iterations adapts an SVDQuant checkpoint.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A closed-form noise-schedule sensitivity score ($S(k) \approx |B_k|/\sqrt{\bar\alpha_{k-1}}$) correlates 0.96–0.97 with TASQ's learned precision, which suggests a pure heuristic schedule could approximate TASQ's allocation with zero training, though the learned masks also capture non-monotonic, layer-specific deviations that a heuristic would miss.
  • TASQ's benefits vanish on hardware whose lowest supported precision is fixed (e.g., INT4 datapaths); the paper's own limitation section concedes this, so the practical win is tied to adoption of precision-scalable or bit-serial execution.
  • Because TASQ adapts only weights and leaves activations fixed, composing it with activation-adaptive schemes like AdaTSQ is a natural next step that could push the Pareto frontier further; the paper lists joint weight–activation precision as future work.
  • The shared-prefix constraint implies that on models where per-stage optimal weights diverge in their MSBs, TASQ would need a small per-stage scale or bias correction to stay competitive; a testable extension would add such corrections and measure the FID gain.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. TASQ proposes a temporal-adaptive weight quantization method for diffusion models. It stores a single high-precision weight buffer and learns a Temporal-Spatial LSB Mask that, per denoising stage and per layer, drops least-significant bit planes, thereby reducing effective weight precision and BitOPs without storing per-stage weight copies. The method is trained with LoRA-based quantization-aware distillation and uses a farthest-stage-first training order. The authors also design a bit-serial Temporal-Precision Engine that streams only the selected weight planes, with cycle counts scaling with effective precision. Experiments on PixArt-Σ, SANA-1.6B, and SDXL-Turbo report FID, ImageReward, LPIPS, and PSNR at W8A8, W4A8, and W4A4 budgets, together with ablations and hardware cycle measurements, claiming quality comparable to static quantization with 25–50% fewer cycles, and up to 7.5× over a naive 8-bit bit-serial execution.

Significance. If the claims hold, TASQ addresses a real and under-explored redundancy in diffusion inference: static weight precision pays worst-case arithmetic at every denoising step. The shared-buffer design with learned LSB masks is a clean and practical mechanism to decouple storage precision from operating precision, and the paper provides a substantial empirical study across three backbones and two datasets. The derivation of the sensitivity schedule in Appendix D.1, and its high correlation (0.97, 0.96) with the learned allocation, is a useful mechanistic interpretation rather than an optimization objective. The hardware co-design is well motivated, and the experiments are reasonably controlled (shared training pools, matched iteration counts). However, the central quality-compute comparison is not memory-matched in the W4 rows, and the paper does not isolate the representational cost of the shared-prefix constraint. These issues affect whether the reported quality gains are attributable to temporal adaptivity or to the extra storage and to the internal representation advantage. With additional baselines and a clarified framing, the contribution would be solid.

major comments (4)
  1. [Section 4, Table 2] The W4A8 and W4A4 comparisons are not memory-matched. TASQ stores an 8-bit shared weight buffer (b_max=8) while the static W4 baselines (SVDQuant, QAT, ViDiT-Q, etc.) store 4-bit weights. At the same BitOPs, TASQ therefore has access to a strictly richer weight representation, so its quality improvements over static W4 could be due to the additional storage rather than to temporal adaptation. The paper does state in Section 4.3 that it does not target storage, but Table 2 presents these rows as direct comparisons. I recommend either (i) adding memory-matched experiments in which b_max equals the static baseline's precision (e.g., 4-bit storage), or (ii) reframing Table 2 to make the storage asymmetry explicit and moving the primary efficiency claim to the W8 rows, where storage is matched and the 25% compute reduction is clean.
  2. [Eqs. (1) and (4), Section 3.2] The core mechanism forces every stage's low-precision weights to be nested prefixes of one common integer (Q_{n-1} = floor(Q_n/2)). The paper provides no baseline that measures the quality cost of this representational constraint. All experiments, including the Table 3 ablation, use the shared-buffer prefix mechanism, so they cannot distinguish the benefit of temporal adaptation from the harm of forcing all stages to share a single high-precision buffer. I recommend adding an upper-bound comparison where each stage is independently re-quantized at its own optimal precision (with per-stage storage), at the same average bit-width. If the gap between TASQ and this per-stage baseline is small, the constraint is empirically justified; if the gap is large, the no-copy advantage should be weighed against a measurable quality loss, which is important for deployment decisions.
  3. [Section 3.2, Eqs. (4)-(8)] The notation and gradient derivation for the LSB mask are ambiguous for multi-bit truncation. Equation (4) defines a single mask M_{t,l} that removes one LSB, but Figure 5 and Section G.2 show that the method removes multiple bit planes (precisions range from 1 to 8 bits). The paper says the operation is applied recursively, but the objective in Eq. (5) sums over M_{t,l} as if it were a single scalar, and Eq. (8) gives a gradient for that single mask. It is unclear whether M_{t,l} is a vector over bit positions, how the recursion is defined, and how the regularization and pruning in Algorithm 1 operate across multiple bits. Please formalize the multi-bit mask, the masked quantizer for k dropped bits, and the corresponding gradients; otherwise the learning rule is not reproducible.
  4. [Section 4.3, Table 6] The paper states that the Temporal-Precision Engine is implemented 'in software based on BISMO' and reports 'measured' cycles and 'no measured cycle overhead' for precision switching. It is not clear whether these are cycle counts from a cycle-accurate simulator, an FPGA configuration, or a software emulation. If they come from a simulation, the simulator and its clock-accuracy model should be described, and the zero-cycle switching claim should be justified at the microarchitectural level (e.g., where the precision register is set and why it does not affect the pipeline). This is needed to evaluate the headline claim of 25–50% cycle reductions.
minor comments (5)
  1. [Section 4.2 / Table 3] The ablation shows that temporal and spatial adaptation are complementary, but it does not include a comparison of the farthest-stage-first training order against sequential or random order. The motivation for this heuristic is described in the text, but a small ablation would make the design choice more convincing.
  2. [Section G.2, Eq. (22)] Equation (22) generalizes the right shift to k bits, but the mask notation M_{t,l} is not defined for multi-bit drops. Please reconcile this with the single-bit definition in Eq. (4) (see major comment 3).
  3. [Section 4.3 / Table 1] The cycle counts for the three dataflow schedules are well approximated by the fetch counts, but the relationship is not exact (e.g., 2LR fetches yields 36,653 cycles for L=8, while the ratio of fetch counts across schedules is only approximately preserved). The paper should state whether cycles include memory latency, shifting, or other overheads that break the simple proportionality.
  4. [Abstract] The abstract says 'reduces execution cycles by 25 to 50 percent over static quantization and by 6.1–7.5x over a naive static 8-bit bit-serial execution.' The two baselines (a static-quantization engine and a naive 8-bit bit-serial loop) are different, which can be confusing. Please clarify the baseline for each number.
  5. [Table 2] Some numerical entries have inconsistent decimal places (e.g., '19.77' under PSNR for SDXL-Turbo W4A4, while other cells use one or two decimals). Please standardize.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: central claims are independently supported; only minor non-load-bearing self-citations are present.

full rationale

I find no circular step in the derivation chain. The load-bearing claims—that bit allocation can vary by denoising stage while storage stays fixed, that quality is preserved at matched BitOPs, and that cycle savings follow from measured bit-serial execution—are not defined into the inputs. Equations (1) and (4) define a shared-buffer truncation scheme, but the per-stage mask values are learned from the distillation loss plus the mask penalty in Eq. (5), not read off from the sensitivity formula; the mask optimization is independent of S(k). Appendix D.1 derives S(k) from the DDIM error-propagation analysis of Liu et al. (2025a) with no fitted constants and then compares it against the learned allocation in Table 10, which is an external consistency check rather than a forced prediction. Hardware cycle counts are measured in Table 6, not fit to the claimed BitOP savings. The self-citations (Han et al. 2025 for LSB regularization and Liu et al. 2025b for annealing) are methodological precedents and are not load-bearing: removing them would not collapse the derivation, because the core QAT objective and hardware measurements stand on their own. The shared-prefix representational constraint is a genuine limitation but not a circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central method introduces no physical entities. The load-bearing assumptions are the truncation-prefix constraint, STE, and the locally-constant LSB approximation, all standard in bit-level QAT but unproved. The free parameters are hyperparameters and the bmax=8 buffer choice, which affects the memory fairness of the W4 comparisons.

free parameters (5)
  • Number of temporal stages (S) = 4
    Chosen via elbow-point analysis on MJHQ (Appendix F); directly sets the granularity of temporal adaptivity and the number of mask parameters.
  • Mask regularization strength (lambda_M) = 5e-5
    Balances distillation loss against LSB retention; fixed across all experiments and settings.
  • Annealing horizon (s0) and temperature schedule = not fully specified in main text
    Controls binarization speed of the continuous gate in Eq. (9)-(10); taken from PAT; values needed to reproduce.
  • LSB pruning interval = 0.1k iterations
    Frequency at which masks are pruned during training (Algorithm 1).
  • Shared buffer width (bmax) = 8 bits (all operating points)
    The paper stores an 8-bit buffer even for W4A4 settings; at W4A4 this doubles weight memory versus a native 4-bit static baseline, so compute comparisons are not memory-matched.
assumptions (5)
  • domain assumption Straight-through estimator (STE) for the floor function in quantization
    Gradients pass through Q_n as if the floor were identity (Eq. 7). Standard in QAT but an approximation.
  • domain assumption Locally constant LSB under weight updates
    Eq. (6) sets dLSB/dW approximately 0 so the mask gradient only sees the LSB value, not its dependence on W; if this fails, the mask training signal is biased.
  • domain assumption Truncation-prefix representational assumption
    Stage-specific operating weights are always prefixes of a single shared quantized integer (Eqs. 1 and 4). The paper gives no proof that this constraint is harmless; it relies on empirical FID results.
  • domain assumption Jacobian-free error propagation (J approximately 0)
    Appendix D.1 derives the timestep sensitivity S(k) using J approximately 0 from Liu et al. 2025a; this underpins the explanatory correlation, not the training itself.
  • domain assumption Cached teacher features as a valid distillation target
    Training uses precomputed teacher features (Appendix A.2) instead of the full teacher, assuming this preserves the quality signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TASQ: Temporal-Adaptive Bit Sparsification Quantization for Diffusion Models." pith.science (2026). https://pith.science/paper/247LU6SX

@misc{pith2026260803057,
  author       = {Pith},
  title        = {Pith review of: TASQ: Temporal-Adaptive Bit Sparsification Quantization for Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/247LU6SX}},
  note         = {Machine review of arXiv:2608.03057}
}
read the original abstract

Static quantization assigns one weight precision to every denoising step. To preserve quality, that precision must accommodate the most quantization-sensitive step, even though many other steps can tolerate fewer bits. The resulting model may satisfy its memory budget, but it repeatedly pays worst-case arithmetic throughout the denoising trajectory. We introduce Temporal-Adaptive Bit Sparsification Quantization (TASQ) to separate these two costs. TASQ stores one shared maximum-precision weight buffer and learns a Temporal-Spatial LSB Mask that selects a lower effective precision for each layer and denoising stage by truncating least-significant bits. Storage therefore remains fixed by the worst case, while BitOPs decrease at less sensitive stages without per-stage weight copies or runtime search. A Temporal-Precision Engine maps the learned schedule to bit-serial execution, where cycles scale with effective precision and switching precision has no measured cycle overhead. On PixArt-Sigma, SANA-1.6B, and SDXL-Turbo, TASQ achieves quality comparable to static quantization with less computation. Together with the Temporal-Precision Engine, it reduces execution cycles by 25 to 50 percent over static quantization and by 6.1 to 7.5x over a naive static 8-bit bit-serial execution. Code is available at https://github.com/seokho-han/tasq.

Figures

Figures reproduced from arXiv: 2608.03057 by the authors.

Figure 1
Figure 1. Temporal sensitivity and TASQ bit allocation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Process of Temporal–Spatial LSB Mask Regularization and LSB Mask Annealing. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative Image Generation Results on PixArt- [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Effective weight precision across temporal stages under the W4A4 operating budget. The static allocation must cover [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: TASQ-learned operating precision across temporal [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Execute cycles per denoising step under the mea [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Per-projection × per-block input-activation RMS, ∥x∥2/ √ N, of SANA-1.6B across denoising steps. cientDM’s IS of 9.41 and lowers FID from 3.80 to 3.65. At A4, it raises IS from 9.37 to 9.40 and lowers FID from 3.91 to 3.90. F Number of Stages [PITH_FULL_IMAGE:figures/…
Figure 8
Figure 8. Figure 8: TASQ-learned operating weight precision across temporal stages (rows) and weight tensors (columns) at the average [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Ablation on the number of temporal stages on MJHQ for (a) SDXL-Turbo and (b) PixArt- [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Qualitative Image Generation Results on SANA-1.6B [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Qualitative Image Generation Results on SDXL-Turbo and PixArt- [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 27 canonical work pages

  1. [1]

    Askarihemmat , M.; Wagner, S.; Bilaniuk, O.; Hariri, Y.; Savaria, Y.; and David, J.-P. 2023. BARVINN : Arbitrary Precision DNN Accelerator Controlled by a RISC-V CPU . In Proceedings of the 28th Asia and South Pacific Design Automation Conference (ASP-DAC)

    work page 2023

  2. [2]

    Chen, J.; Ge, C.; Xie, E.; Wu, Y.; Yao, L.; Ren, X.; Wang, Z.; Luo, P.; Lu, H.; and Li, Z. 2024. PixArt-Sigma: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation. arXiv:2403.04692

    arXiv 2024

  3. [3]

    Chen, L.; Meng, Y.; Tang, C.; Ma, X.; Jiang, J.; Wang, X.; Wang, Z.; and Zhu, W. 2025. Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 28306--28315

    work page 2025

  4. [4]

    Chen, X.; Fang, H.; Lin, T.-Y.; Vedantam, R.; Gupta, S.; Dollar, P.; and Zitnick, C. L. 2015. Microsoft COCO Captions: Data Collection and Evaluation Server. arXiv preprint arXiv:1504.00325

    arXiv 2015

  5. [5]

    Dhariwal, P.; and Nichol, A. 2021. Diffusion models beat GANs on image synthesis. In Proceedings of the 35th International Conference on Neural Information Processing Systems, NIPS '21. Red Hook, NY, USA: Curran Associates Inc. ISBN 9781713845393

    work page 2021

  6. [6]

    W.; and Keutzer, K

    Dong, Z.; Yao, Z.; Arfeen, D.; Gholami, A.; Mahoney, M. W.; and Keutzer, K. 2020. HAWQ-V2: Hessian Aware Trace-Weighted Quantization of Neural Networks. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, 18518--18529

    work page 2020

  7. [7]

    W.; and Keutzer, K

    Dong, Z.; Yao, Z.; Gholami, A.; Mahoney, M. W.; and Keutzer, K. 2019. HAWQ: Hessian Aware Quantization of Neural Networks with Mixed-Precision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 293--302

    work page 2019

  8. [8]

    Feng, W.; Qin, H.; Yang, C.; An, Z.; Huang, L.; Diao, B.; Wang, F.; Tao, R.; Xu, Y.; and Magno, M. 2025. MPQ-DM : Mixed Precision Quantization for Extremely Low Bit Diffusion Models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)

    work page 2025

Show all 46 references
  1. [9]

    W.; and Keutzer, K

    Gholami, A.; Kim, S.; Dong, Z.; Yao, Z.; Mahoney, M. W.; and Keutzer, K. 2021. A Survey of Quantization Methods for Efficient Neural Network Inference. arXiv preprint arXiv:2103.13630

    2021 arXiv

  2. [10]

    E.; Yang, H.; and Ko, J

    Han, S.; Yoon, S.; Kim, J.; Wang, D.; Jeon, K. E.; Yang, H.; and Ko, J. H. 2025. MSQ: Memory-Efficient Bit Sparsification Quantization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 21885--21894

    2025

  3. [11]

    He, Y.; Liu, J.; Wu, W.; Zhou, H.; and Zhuang, B. 2024. EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models. In International Conference on Learning Representations (ICLR)

    2024

  4. [12]

    He, Y.; Liu, L.; Liu, J.; Wu, W.; Zhou, H.; and Zhuang, B. 2023. PTQD: Accurate Post-Training Quantization for Diffusion Models. arXiv preprint arXiv:2305.10657

    2023 arXiv

  5. [13]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20. Red Hook, NY, USA: Curran Associates Inc. ISBN 9781713829546

    2020

  6. [14]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2022. Lo RA : Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations

    2022

  7. [15]

    Huang, H.; Chen, J.; Guo, J.; Zhan, R.; and Wang, Y. 2025. TCAQ-DM : Timestep-Channel Adaptive Quantization for Diffusion Models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)

    2025

  8. [16]

    Huang, Y.; Gong, R.; Liu, J.; Chen, T.; and Liu, X. 2024. TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2024

  9. [17]

    Jeon, H.; Kim, Y.; and Kim, J.-j. 2025. L4Q: Parameter Efficient Quantization-Aware Fine-Tuning on Large Language Models. arXiv:2402.04902

    2025 arXiv

  10. [18]

    Li, D.; Kamko, A.; Akhgari, E.; Sabet, A.; Xu, L.; and Doshi, S. 2024. Playground v2.5: Three Insights towards Enhancing Aesthetic Quality in Text-to-Image Generation. arXiv:2402.17245

    2024 arXiv

  11. [19]

    Li, M.; Lin, Y.; Zhang, Z.; Cai, T.; Li, X.; Guo, J.; Xie, E.; Meng, C.; Zhu, J.-Y.; and Han, S. 2025. SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models. In The Thirteenth International Conference on Learning Representations

    2025

  12. [20]

    Li, X.; Liu, Y.; Lian, L.; Yang, H.; Dong, Z.; Kang, D.; Zhang, S.; and Keutzer, K. 2023. Q-Diffusion: Quantizing Diffusion Models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 17535--17545

    2023

  13. [21]

    Liu, S.; Zeng, C.; Yan, C.; Peng, X.; Wang, X.; Chen, F.; and Mei, X. 2025 a . Error Propagation Mechanisms and Compensation Strategies for Quantized Diffusion. arXiv preprint arXiv:2508.12094

    2025

  14. [22]

    Liu, Y.; Yang, H.; Chen, Y.; Zhang, R.; Wang, M.; Du, Y.; and Du, L. 2025 b . PAT: Pruning-Aware Tuning for Large Language Models. Proceedings of the AAAI Conference on Artificial Intelligence, 39(23): 24686--24695

    2025

  15. [23]

    M.; Lewandowski, B.; and Chen, L

    Maruzzelli, R. M.; Lewandowski, B.; and Chen, L. Y. 2024. MPQ-Diff : Mixed Precision Quantization for Diffusion Models. arXiv:2412.00144

    2024 arXiv

  16. [24]

    Podell, D.; English, Z.; Lacey, K.; Blattmann, A.; Dockhorn, T.; Muller, J.; Penna, J.; and Rombach, R. 2023. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. ArXiv, abs/2307.01952

    2023 arXiv

  17. [25]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10684--10695

    2022

  18. [26]

    Sauer, A.; Lorenz, D.; Blattmann, A.; and Rombach, R. 2023. Adversarial Diffusion Distillation. arXiv preprint arXiv:2311.17042

    2023 arXiv

  19. [27]

    Shang, Y.; Yuan, Z.; Xie, B.; Wu, B.; and Yan, Y. 2023. Post-Training Quantization on Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1972--1981

    2023

  20. [28]

    Sui, Y.; Li, Y.; Kag, A.; Idelbayev, Y.; Cao, J.; Hu, J.; Sagar, D.; Yuan, B.; Tulyakov, S.; and Ren, J. 2024. BitsFusion: 1.99 bits Weight Quantization of Diffusion Model. In Globerson, A.; Mackey, L.; Belgrave, D.; Fan, A.; Paquet, U.; Tomczak, J.; and Zhang, C., eds., Advan...

    2024

  21. [29]

    Tang, S.; Wang, X.; Chen, H.; Guan, C.; Wu, Z.; Tang, Y.; and Zhu, W. 2024. Post-Training Quantization with Progressive Calibration and Activation Relaxing for Text-to-Image Diffusion Models. In Proceedings of the European Conference on Computer Vision (ECCV)

    2024

  22. [30]

    Umuroglu, Y.; Rasnayake, L.; and Sj \"a lander, M. 2018. BISMO : A Scalable Bit-Serial Matrix Multiplication Overlay for Reconfigurable Computing. In 28th International Conference on Field Programmable Logic and Applications (FPL), 307--314

    2018

  23. [31]

    Urbanek, J.; Bordes, F.; Astolfi, P.; Williamson, M.; Sharma, V.; and Romero-Soriano, A. 2024. A Picture is Worth More Than 77 Text Tokens: Evaluating CLIP-Style Models on Dense Captions. arXiv:2312.08578

    2024 arXiv

  24. [32]

    Wang, C.; Peng, H.-Y.; Liu, Y.-T.; Gu, J.; and Hu, S.-M. 2025 a . Diffusion Models for 3D Generation: A Survey. Comput. Vis. Media, 11(1): 1--28

    2025

  25. [33]

    Wang, C.; Wang, Z.; Xu, X.; Tang, Y.; Zhou, J.; and Lu, J. 2024. Towards Accurate Post-training Quantization for Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 16026--16035

    2024

  26. [34]

    Wang, H.; Shang, Y.; Yuan, Z.; Wu, J.; Yan, J.; and Yan, Y. 2025 b . QuEST: Low-bit Diffusion Model Quantization via Efficient Selective Finetuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 15542--15551

    2025

  27. [35]

    Wang, K.; Liu, Z.; Lin, Y.; Lin, J.; and Han, S. 2019. HAQ: Hardware-Aware Automated Quantization with Mixed Precision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 8612--8620

    2019

  28. [36]

    Wang, K.; Shi, M.; Zhou, Y.; Li, Z.; Yuan, Z.; Shang, Y.; Peng, X.; Zhang, H.; and You, Y. 2025 c . A Closer Look at Time Steps is Worthy of Triple Speed-Up for Diffusion Model Training. arXiv preprint arXiv:2405.17403

    2025 arXiv

  29. [37]

    Whalen, L.; Du, Z.; You, H.; Li, C.; Li, S.; and Lin, Y. 2025. Early-Bird Diffusion: Investigating and Leveraging Timestep-Aware Early-Bird Tickets in Diffusion Models for Efficient Training. arXiv preprint arXiv:2504.09606

    2025 arXiv

  30. [38]

    Wu, J.; Wang, H.; Shang, Y.; Shah, M.; and Yan, Y. 2024. PTQ4DiT: Post-training Quantization for Diffusion Transformers. In Globerson, A.; Mackey, L.; Belgrave, D.; Fan, A.; Paquet, U.; Tomczak, J.; and Zhang, C., eds., Advances in Neural Information Processing Systems, volume...

    2024

  31. [39]

    Xiao, L.; Yang, H.; Dong, Z.; Keutzer, K.; Du, L.; and Zhang, S. 2023. Csq: Growing mixed-precision quantization scheme with bi-level continuous sparsification. In 2023 60th ACM/IEEE Design Automation Conference (DAC), 1--6. IEEE

    2023

  32. [40]

    Xie, E.; Chen, J.; Chen, J.; Cai, H.; Tang, H.; Lin, Y.; Zhang, Z.; Li, M.; Zhu, L.; Lu, Y.; and Han, S. 2024. SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers. arXiv preprint arXiv:2410.10629

    2024 arXiv

  33. [41]

    Yang, H.; Duan, L.; Chen, Y.; and Li, H. 2021. BSQ: Exploring bit-level sparsity for mixed-precision neural network quantization. arXiv preprint arXiv:2102.10462

    2021 arXiv

  34. [42]

    Yang, Y.; Dai, X.; Wang, J.; Zhang, P.; and Zhang, H. 2023. Efficient Quantization Strategies for Latent Diffusion Models. arXiv preprint arXiv:2312.05431

    2023 arXiv

  35. [43]

    Zhang, S.; Ding, Z.; Yang, K.; Wu, J.; Yan, X.; Li, X.; Duan, B.; Fang, J.; and Zhang, Y. 2026. AdaTSQ : Pushing the Pareto Frontier of Diffusion Transformers via Temporal-Sensitivity Quantization. arXiv:2602.09883

    2026

  36. [44]

    Zhao, T.; Fang, T.; Huang, H.; Liu, E.; Wan, R.; Soedarmadji, W.; Li, S.; Lin, Z.; Dai, G.; Yan, S.; Yang, H.; Ning, X.; and Wang, Y. 2024 a . ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation. arXiv preprint arXiv:2406.02540

    2024 arXiv

  37. [45]

    Zhao, T.; Ning, X.; Fang, T.; Liu, E.; Huang, G.; Lin, Z.; Yan, S.; Dai, G.; and Wang, Y. 2024 b . MixDQ: Memory-Efficient Few-Step Text-to-Image Diffusion Models with Metric-Decoupled Mixed Precision Quantization. In Proceedings of the European Conference on Computer Vision (ECCV)

    2024

  38. [46]

    Zheng, X.; Qin, H.; Ma, X.; Zhang, M.; Hao, H.; Wang, J.; Zhao, Z.; Guo, J.; and Liu, X. 2024. BinaryDM: Towards Accurate Binarization of Diffusion Model. arXiv preprint arXiv:2404.05662

    2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.