Pith. sign in

REVIEW 3 major objections 1 minor 40 references

HeatKV doubles KV-cache compression in visual autoregressive models by ranking heads on attention to prior scales and applying a static pruning schedule.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-30 21:22 UTC pith:VNFUHXCL

load-bearing objection HeatKV gives a static per-head pruning schedule for VAR KV caches that claims 2x compression on Infinity-2B with no quality drop, but the experiments and generalization checks are too thin to support the SOTA claim yet. the 3 major comments →

arxiv 2605.14877 v2 pith:VNFUHXCL submitted 2026-05-14 cs.CV

HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling

Jonathan Cederlund , Axel Berg , William Isaksson , Durmus Alp Emre Acar , Chuteng Zhou , Pontus Giselsson This is my paper
classification cs.CV
keywords KV cache compressionvisual autoregressive modelsattention head rankingimage generationmemory efficiencypruning schedule
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Visual autoregressive models generate images one scale at a time yet require gigabytes of KV cache per image. HeatKV ranks each attention head by its attention scores over earlier scales using a small offline calibration set, then builds one fixed pruning schedule that removes more cache from lower-ranked heads. Applied to the Infinity-2B model, this head-specific allocation reaches twice the compression ratio of prior methods. Image fidelity, prompt alignment, and human perception scores remain the same or better, setting a new state of the art for VAR KV-cache compression.

Core claim

Ranking attention heads according to their attention scores over prior scales from a small offline calibration set produces a static pruning schedule that supports fine-grained, head-specific cache allocation; when applied to the Infinity-2B model under a fixed memory budget, this schedule delivers a 2 imes higher compression ratio than existing methods while maintaining or improving image fidelity, prompt alignment, and human perception scores.

What carries the argument

Static pruning schedule derived from ranking heads by attention scores over prior scales, enabling head-specific KV cache allocation.

Load-bearing premise

The ranking of attention heads by their attention scores over prior scales, obtained from a small offline calibration set, produces a static pruning schedule that generalizes reliably to arbitrary prompts and generated images without degrading quality.

What would settle it

Apply the same pruning schedule to a large set of prompts outside the calibration data and measure whether image quality metrics fall below the uncompressed baseline.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Twice the KV-cache compression ratio is achieved compared with existing methods at matched or higher quality.
  • New state-of-the-art compression performance is reached for VAR models.
  • Head-specific allocation outperforms uniform pruning under the same memory budget.
  • Quality metrics (fidelity, prompt alignment, human perception) stay comparable or improve.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The calibration-derived ranking may transfer to other autoregressive image or video models if their attention patterns across scales are similar.
  • Static schedules could be paired with lightweight runtime adjustments to handle outlier prompts.
  • Reduced memory footprint may allow higher-resolution generation or larger batch sizes on fixed hardware.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 1 minor

Summary. The manuscript introduces HeatKV, a KV-cache compression technique for Visual Autoregressive (VAR) models. Attention heads are ranked by their attention scores over prior scales on a small offline calibration set; a static per-head pruning schedule is then derived for any target memory budget. On the Infinity-2B model the method is reported to deliver a 2× higher compression ratio than existing approaches while preserving or improving image fidelity, prompt alignment, and human perception scores, establishing a new SOTA for VAR KV-cache compression.

Significance. If the reported gains prove robust, the head-specific static schedule offers a practical route to substantially lower memory footprints for large VAR models without retraining, which would be valuable for deployment. The fine-grained allocation is a clear departure from uniform or layer-wise baselines.

major comments (3)
  1. [Abstract] Abstract: the central empirical claim of 2× higher compression and SOTA status is asserted without any description of the evaluation protocol, baseline implementations, number of samples, statistical tests, or controls for selection effects, leaving the soundness of the result only moderately supported.
  2. [Method description] The method constructs a static pruning schedule from attention scores on a small calibration set and asserts that this schedule generalizes to arbitrary prompts while maintaining 'similar or better' quality; however, the manuscript supplies no information on calibration-set size, diversity, or explicit hold-out testing on unseen inputs, which is load-bearing for the generalization guarantee.
  3. [Method description] The per-head allocation percentages are free parameters fitted on the calibration set; the manuscript does not demonstrate that the resulting schedule remains effective when the calibration distribution differs from the test distribution, undermining the claim that the schedule is reliably 'tailored to a given memory budget'.
minor comments (1)
  1. The GitHub link for code and calibration script is helpful for reproducibility but should be accompanied by a brief description of the calibration-set construction in the main text.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the constructive feedback. We address each major comment below and will revise the manuscript accordingly to improve clarity on evaluation details and generalization aspects.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central empirical claim of 2× higher compression and SOTA status is asserted without any description of the evaluation protocol, baseline implementations, number of samples, statistical tests, or controls for selection effects, leaving the soundness of the result only moderately supported.

    Authors: We agree the abstract would benefit from more context on the evaluation. In revision we will expand it to note that results use the Infinity-2B model on standard image-generation benchmarks, comparing against prior KV-cache methods with metrics of image fidelity, prompt alignment, and human perception scores. The full protocol (including sample counts, baseline implementations, and identical test conditions for all methods) appears in the experimental section; no statistical significance tests were run because gains were uniform across prompts, but we will add a sentence clarifying this and the absence of selection effects. revision: yes

  2. Referee: [Method description] The method constructs a static pruning schedule from attention scores on a small calibration set and asserts that this schedule generalizes to arbitrary prompts while maintaining 'similar or better' quality; however, the manuscript supplies no information on calibration-set size, diversity, or explicit hold-out testing on unseen inputs, which is load-bearing for the generalization guarantee.

    Authors: We acknowledge that explicit details on calibration-set size, diversity, and hold-out validation should be stated in the main text rather than only the code release. In revision we will add a dedicated paragraph specifying the calibration set (size, source, and diversity criteria) and report that the derived schedule was evaluated on a held-out prompt set disjoint from calibration, confirming stable quality. This directly supports the generalization statement. revision: yes

  3. Referee: [Method description] The per-head allocation percentages are free parameters fitted on the calibration set; the manuscript does not demonstrate that the resulting schedule remains effective when the calibration distribution differs from the test distribution, undermining the claim that the schedule is reliably 'tailored to a given memory budget'.

    Authors: The allocations are indeed derived from calibration rankings, and the paper relies on the empirical stability of head-wise attention patterns across prompts in VAR's scale-based generation. To address the distribution-shift concern we will add a short robustness experiment in the revision, re-deriving the schedule from an alternate calibration distribution and verifying comparable performance on the original test set. This will strengthen the claim that the schedule is reliably tailored. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical calibration method is self-contained

full rationale

The paper presents HeatKV as an empirical procedure: rank heads by attention scores on a small offline calibration set, then build a static pruning schedule for a target memory budget. No equations, derivations, or self-citations reduce the reported 2× compression ratio, fidelity metrics, or SOTA claim back to inputs by construction. The calibration set is external data; the schedule is a fixed output of that process, not a fitted parameter redefined as a prediction. No self-citation load-bearing, uniqueness theorems, or ansatzes smuggled via prior work appear in the provided text. The central result is an observed performance number on Infinity-2B, which stands independently of the listed circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach relies on one main domain assumption about generalization of calibration rankings and introduces a data-derived pruning schedule as its primary fitted element.

free parameters (1)
  • per-head allocation percentages in the pruning schedule
    Chosen from calibration-set attention rankings to meet the target memory budget
axioms (1)
  • domain assumption Attention head importance measured on the calibration set remains stable and predictive for all future generations
    This premise justifies building a single static schedule instead of dynamic per-prompt pruning

pith-pipeline@v0.9.1-grok · 5724 in / 1230 out tokens · 49432 ms · 2026-06-30T21:22:18.983440+00:00 · methodology

0 comments
read the original abstract

Visual Autoregressive (VAR) models have recently demonstrated impressive image generation quality while maintaining low latency. However, they suffer from severe KV-cache memory constraints, often requiring gigabytes of memory per generated image. We introduce HeatKV, a novel compression method that adapts cache allocation in each head based on its attention to previously generated scales. Using a small offline calibration set, the attention heads are ranked according to their attention scores over prior scales. Based on this ranking, we construct a static pruning schedule tailored to a given memory budget. Applied to the Infinity-2B model, HeatKV achieves $2 \times$ higher compression ratio in memory allocation for KV cache compared to existing methods, while maintaining similar or better image fidelity, prompt alignment and human perception score. Our method achieves a new state-of-the-art (SOTA) for VAR model KV-cache compression, showcasing the effectiveness of fine-grained, head-specific cache allocation. Code and calibration script available at https://github.com/arm-research/heatkv.

Figures

Figures reproduced from arXiv: 2605.14877 by Axel Berg, Chuteng Zhou, Durmus Alp Emre Acar, Jonathan Cederlund, Pontus Giselsson, William Isaksson.

Figure 1
Figure 1. Figure 1: Example images generated with Infinity-8B [1] using HeatKV at [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of HeatKV. (1) VARs generate image in increasing scale size, causing growth [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Different head-attention patterns. The y-axis shows tokens in the current scale (raster [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Removal order using BINARY-HEATKV with GREEDYEARLYPRUNING in Algorithm 1 on Infinity-2B, 10% budget. Each small square is an attention head, specified by layer and head index. ranking head from least to most dependent on tokens from i. Following Algorithm 1, after scale k we form Gk by taking the prefix of length Nk from each Oi , (i ≤ k) and uniting them. Thus, Gk is the set that must be absent by the end… view at source ↗
Figure 5
Figure 5. Figure 5: Infinity-2B results on MS-COCO 2017 and H100 generation speed using 10% budget. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Stability of S-CAS head-scale rankings across calibration runs for Infinity-2B and Infinity [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Infinity-2B [1] with HeatKV at 10% budget [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Infinity-8B [1] with HeatKV at 10% budget [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 4 canonical work pages · 2 internal anchors

  1. [1]

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis,

    J. Han, J. Liu, Y . Jiang, B. Yan, Y . Zhang, Z. Yuan, B. Peng, and X. Liu, “Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis,” inProceedings of the Computer Vision and Pattern Recognition Conference, pp. 15733–15744, 2025

    2025

  2. [2]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,”Advances in neural information processing systems, vol. 27, 2014

    2014

  3. [3]

    A style-based generator architecture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 4401–4410, 2019

    2019

  4. [4]

    Stylegan-xl: Scaling stylegan to large diverse datasets,

    A. Sauer, K. Schwarz, and A. Geiger, “Stylegan-xl: Scaling stylegan to large diverse datasets,” inACM SIGGRAPH 2022 conference proceedings, pp. 1–10, 2022

    2022

  5. [5]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural informa- tion processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  6. [6]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10684–10695, 2022

    2022

  7. [7]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 4195–4205, 2023

    2023

  8. [8]

    Visual autoregressive modeling: Scalable image gener- ation via next-scale prediction,

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, and L. Wang, “Visual autoregressive modeling: Scalable image gener- ation via next-scale prediction,”Advances in neural information processing systems, vol. 37, pp. 84839– 84865, 2024

    2024

  9. [9]

    Memory-efficient visual autoregressive modeling with scale-aware kv cache compression,

    K. Li, Z. Chen, C.-Y . Yang, and J.-N. Hwang, “Memory-efficient visual autoregressive modeling with scale-aware kv cache compression,”Advances in Neural Information Processing Systems, 2025

    2025

  10. [10]

    Generative pretraining from pixels,

    M. Chen, A. Radford, R. Child, J. Wu, H. Jun, D. Luan, and I. Sutskever, “Generative pretraining from pixels,” inInternational conference on machine learning, pp. 1691–1703, PMLR, 2020

    2020

  11. [11]

    Taming transformers for high-resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high-resolution image synthesis,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 12873–12883, 2021

    2021

  12. [12]

    Scaling autoregressive models for content-rich text-to-image generation,

    J. Yu, Y . Xu, J. Y . Koh, T. Luong, G. Baid, Z. Wang, V . Vasudevan, A. Ku, Y . Yang, B. K. Ayan, B. Hutchinson, W. Han, Z. Parekh, X. Li, H. Zhang, J. Baldridge, and Y . Wu, “Scaling autoregressive models for content-rich text-to-image generation,”Transactions on Machine Learning Research, 2022

    2022

  13. [13]

    Parallelized autoregres- sive visual generation,

    Y . Wang, S. Ren, Z. Lin, Y . Han, H. Guo, Z. Yang, D. Zou, J. Feng, and X. Liu, “Parallelized autoregres- sive visual generation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12955–12965, 2025

    2025

  14. [14]

    Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation

    P. Sun, Y . Jiang, S. Chen, S. Zhang, B. Peng, P. Luo, and Z. Yuan, “Autoregressive model beats diffusion: Llama for scalable image generation,”arXiv preprint arXiv:2406.06525, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    HART: Efficient visual generation with hybrid autoregressive transformer,

    H. Tang, Y . Wu, S. Yang, E. Xie, J. Chen, J. Chen, Z. Zhang, H. Cai, Y . Lu, and S. Han, “HART: Efficient visual generation with hybrid autoregressive transformer,” inThe Thirteenth International Conference on Learning Representations, 2025. 10

    2025

  16. [16]

    FlowAR: Scale-wise autoregressive image generation meets flow matching,

    S. Ren, Q. Yu, J. He, X. Shen, A. Yuille, and L.-C. Chen, “FlowAR: Scale-wise autoregressive image generation meets flow matching,” inForty-second International Conference on Machine Learning, 2025

    2025

  17. [17]

    FlexV AR: Flexible visual autoregressive modeling without residual prediction,

    S. Jiao, G. Zhang, Y . Qian, J. Huang, Y . Zhao, H. Shi, L. Ma, Y . Wei, and Z. JIE, “FlexV AR: Flexible visual autoregressive modeling without residual prediction,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026

    2026

  18. [18]

    Visual autoregressive modelling for monocular depth estimation,

    A. El-Ghoussani, A. Kaup, N. Navab, G. Carneiro, and V . Belagiannis, “Visual autoregressive modelling for monocular depth estimation,”arXiv preprint arXiv:2512.22653, 2025

    work page arXiv 2025

  19. [19]

    Visual autoregressive modeling for image super-resolution,

    Y . Qu, K. Yuan, J. Hao, K. Zhao, Q. Xie, M. Sun, and C. Zhou, “Visual autoregressive modeling for image super-resolution,” inForty-second International Conference on Machine Learning, 2025

    2025

  20. [20]

    Infinitystar: Unified spacetime autoregressive modeling for visual generation,

    J. Liu, J. Han, B. Yan, Wuhui, F. Zhu, X. Wang, Y . Jiang, B. PENG, and Z. Yuan, “Infinitystar: Unified spacetime autoregressive modeling for visual generation,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026

    2026

  21. [21]

    Efficient streaming language models with attention sinks,

    G. Xiao, Y . Tian, B. Chen, S. Han, and M. Lewis, “Efficient streaming language models with attention sinks,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  22. [22]

    H2o: Heavy-hitter oracle for efficient generative inference of large language models,

    Z. Zhang, Y . Sheng, T. Zhou, T. Chen, L. Zheng, R. Cai, Z. Song, Y . Tian, C. Ré, C. Barrett,et al., “H2o: Heavy-hitter oracle for efficient generative inference of large language models,” 2023

    2023

  23. [23]

    Snapkv: Llm knows what you are looking for before generation,

    Y . Li, Y . Huang, B. Yang, B. Venkitesh, A. Locatelli, H. Ye, T. Cai, P. Lewis, and D. Chen, “Snapkv: Llm knows what you are looking for before generation,” 2024

    2024

  24. [24]

    Ada-KV: Optimizing KV cache eviction by adaptive budget allocation for efficient LLM inference,

    Y . Feng, J. Lv, Y . Cao, X. Xie, and S. K. Zhou, “Ada-KV: Optimizing KV cache eviction by adaptive budget allocation for efficient LLM inference,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  25. [25]

    Head-aware kv cache compression for efficient visual autoregressive modeling,

    Z. Qin, Y . Lv, M. Lin, H. Guo, Z. Zhang, D. Zou, and W. Lin, “Head-aware kv cache compression for efficient visual autoregressive modeling,” 2026

    2026

  26. [26]

    Ams-kv: Adaptive kv caching in multi-scale visual autoregressive transformers,

    B. Xu, Y . Wang, Z. Wang, and P. Li, “Ams-kv: Adaptive kv caching in multi-scale visual autoregressive transformers,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, pp. 27206–27214, 2026

    2026

  27. [27]

    Neural discrete representation learning,

    A. Van Den Oord, O. Vinyals,et al., “Neural discrete representation learning,”Advances in neural information processing systems, vol. 30, 2017

    2017

  28. [28]

    Language model beats diffusion - tokenizer is key to visual generation,

    L. Yu, J. Lezama, N. B. Gundavarapu, L. Versari, K. Sohn, D. Minnen, Y . Cheng, A. Gupta, X. Gu, A. G. Hauptmann, B. Gong, M.-H. Yang, I. Essa, D. A. Ross, and L. Jiang, “Language model beats diffusion - tokenizer is key to visual generation,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  29. [29]

    Image and video tokenization with binary spherical quantization,

    Y . Zhao, Y . Xiong, and P. Kraehenbuehl, “Image and video tokenization with binary spherical quantization,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  30. [30]

    Qwen3 technical report,

    Q. Team, “Qwen3 technical report,” 2025

    2025

  31. [31]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft coco: Common objects in context,” 2014

    2014

  32. [32]

    The unreasonable effectiveness of deep fea- tures as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep fea- tures as a perceptual metric,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 586–595, 2018

    2018

  33. [33]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” inNeural Information Processing Systems, 2017

    2017

  34. [34]

    Geneval: An object-focused framework for evaluating text- to-image alignment,

    D. Ghosh, H. Hajishirzi, and L. Schmidt, “Geneval: An object-focused framework for evaluating text- to-image alignment,” inAdvances in Neural Information Processing Systems, vol. 36, pp. 52132–52152, 2023

    2023

  35. [35]

    Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis

    X. Wu, Y . Hao, K. Sun, Y . Chen, F. Zhu, R. Zhao, and H. Li, “Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis,”arXiv preprint arXiv:2306.09341, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  36. [36]

    Flashattention-2: Faster attention with better parallelism and work partitioning,

    T. Dao, “Flashattention-2: Faster attention with better parallelism and work partitioning,” inThe Twelfth International Conference on Learning Representations, 2023

    2023

  37. [37]

    Fastvar: Linear visual autoregressive modeling via cached token pruning,

    H. Guo, Y . Li, T. Zhang, J. Wang, T. Dai, S.-T. Xia, and L. Benini, “Fastvar: Linear visual autoregressive modeling via cached token pruning,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 19011–19021, 2025

    2025

  38. [38]

    Sparvar: Exploring sparsity in visual autoregressive modeling for training-free acceleration,

    Z. Li, N. Wang, T. Bai, C. Mei, P. Wang, S. Qiu, and J. Cheng, “Sparvar: Exploring sparsity in visual autoregressive modeling for training-free acceleration,”arXiv preprint arXiv:2602.04361, 2026. 11

    work page arXiv 2026

  39. [39]

    Litevar: Compressing visual autoregressive modelling with efficient attention and quantization,

    R. Xie, T. Zhao, Z. Yuan, R. Wan, W. Gao, Z. Zhu, X. Ning, and Y . Wang, “Litevar: Compressing visual autoregressive modelling with efficient attention and quantization,” inWorkshop on Machine Learning and Compression, NeurIPS 2024, 2024

    2024

  40. [40]

    PTQ4ARVG: Post-training quantization for autoregres- sive visual generation models,

    X. Liu, Z. Li, J. Zhang, M. Chen, J. Li, and Q. Gu, “PTQ4ARVG: Post-training quantization for autoregres- sive visual generation models,” inThe Fourteenth International Conference on Learning Representations, 2026. 12 A Additional algorithms Algorithm 2CACHESIZEAFTERLAYER(k, ℓ, G k, Gk−1, Ek) Require: scale k, current layer ℓ, target pruning set Gk, previ...

    2026