pith. sign in

arxiv: 2605.18736 · v2 · pith:KLWO234Znew · submitted 2026-05-18 · 💻 cs.CV

Spectral Progressive Diffusion for Efficient Image and Video Generation

Howard Xiao , Brian Chao , Lior Yariv , Gordon Wetzstein This is my paper

Pith reviewed 2026-05-21 07:48 UTC · model grok-4.3

classification 💻 cs.CV
keywords diffusion modelsimage generationvideo generationefficient inferencefrequency domainprogressive resolutionspectral noise
0
0 comments X

The pith

Diffusion models generate images and videos faster by starting at low resolution and growing it as denoising proceeds from low to high frequencies.

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

The paper establishes that diffusion models implicitly build visual content from low frequencies early in denoising to high frequencies later. This ordering makes high-resolution computation wasteful in the initial noisy stages. Spectral Progressive Diffusion exploits the pattern by expanding resolution along the trajectory using a spectral noise mechanism and a schedule taken from the model's power spectrum. The approach works on existing pretrained models either without any training or with a lightweight fine-tuning step, delivering measurable speedups on both image and video generators while visual quality remains intact.

Core claim

Diffusion models generate visual content autoregressively in the frequency domain, with low-frequency components appearing earlier in the denoising process and high-frequency details emerging later. High-resolution computation on noise-dominated frequencies is therefore redundant. Spectral Progressive Diffusion progressively grows resolution along the denoising trajectory of pretrained models by means of a spectral noise expansion mechanism and an optimal resolution schedule derived from the model's power spectrum. This framework supports both training-free acceleration and a fine-tuning recipe that further improves efficiency and quality.

What carries the argument

Spectral noise expansion mechanism that progressively grows resolution along the denoising trajectory according to a schedule derived from the model's power spectrum.

If this is right

  • Pretrained image and video diffusion models can be accelerated without retraining.
  • A lightweight fine-tuning stage yields further gains in speed and output quality.
  • The same frequency-progression logic applies across both static images and temporal video sequences.
  • Compute savings scale with the length of the denoising trajectory and the chosen resolution schedule.

Where Pith is reading between the lines

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

  • The method could be combined with existing sampler accelerations such as fewer steps or distillation to compound speedups.
  • Similar progressive schedules might transfer to other generative paradigms that exhibit frequency ordering during synthesis.
  • Real-time or edge-device deployment becomes more feasible once early low-resolution stages replace full-resolution passes.

Load-bearing premise

High-resolution computation on noise-dominated frequencies is largely redundant.

What would settle it

Running the identical pretrained model at full resolution throughout denoising produces images or videos of equal or higher quality in equal or less wall-clock time than the progressive schedule.

Figures

Figures reproduced from arXiv: 2605.18736 by Brian Chao, Gordon Wetzstein, Howard Xiao, Lior Yariv.

Figure 1
Figure 1. Figure 1: Spectral Progressive Diffusion. We progressively grow the resolution along the denoising trajectory using an optimal resolution schedule derived from the spectral power of pretrained models (left). At each scheduled transition, our spectral noise expansion mechanism (right) injects high￾frequency noise at the correct level while preserving the partially-denoised low-frequency content. denoised low-frequenc… view at source ↗
Figure 2
Figure 2. Figure 2: Diffusion process in the spectral domain. Latent power spectra in both image and video models decay rapidly with frequency (Fig. (a)), consistent with natural images. Diffusion exhibits a frequency-domain autoregressive structure (Fig. (b)) due to the aforementioned property: low frequencies emerge early in the denoising process, while high frequencies remain noise-dominated. 4 Spectral Progressive Diffusi… view at source ↗
Figure 3
Figure 3. Figure 3: Visual Generation Qualitative Comparisons. For the main comparison of latent-space image generation, our method outperforms the state-of-the-art spatial acceleration method RALU [32] in both visual fidelity and inference speed. Across all evaluated modalities (latent/pixel-space image generation and latent-space video generation), we achieve substantial acceleration over standard high-resolution baselines … view at source ↗
Figure 4
Figure 4. Figure 4: (a): Ablation Studies on δ, S and TΦ. We observe a clear tradeoff between image quality and efficiency when varying δ and S as shown in the top plot. Across transforms, DCT achieves similar quality as DWT and outperforms FFT as shown in the bottom plot. (b): Frequency-based Image Editing. Our method demonstrates superior prompt alignment and geometric consistency compared to standard SDEdit-style spatial-d… view at source ↗
Figure 5
Figure 5. Figure 5: Spectral noise passthrough experiment. At smaller δ ∈ [0.0001, 0.001], there is almost no observable difference compared to native full-resolution generation. As larger δ values cause high-frequency replacement to persist later in the denoising trajectory, we observe increasingly blurry and distorted results (i.e., ghosting artifacts and “CHOOLBUS” instead of “SCHOOLBUS”). 23 [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparisons on latent-space image generation. We compare our method against default-step generation, reduced-step native-resolution generation on FLUX.1-dev [39], and RALU [32], a state-of-the-art acceleration baseline matched to similar speedups. Our method outperforms both baselines. FLUX.1-dev with reduced steps degrades image quality and exhibits over-saturation artifacts, while RALU introd… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparisons on latent-space image generation. We compare our method against default-step generation, reduced-step native-resolution generation on FLUX.1-dev [39], and RALU [32], a state-of-the-art acceleration baseline matched to similar speedups. Our method outperforms both baselines. FLUX.1-dev with reduced steps degrades image quality and exhibits over-saturation artifacts, while RALU introd… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparisons on latent-space image generation. We compare our method against default-step generation, reduced-step native-resolution generation on FLUX.1-dev [39], and RALU [32], a state-of-the-art acceleration baseline matched to similar speedups. Our method outperforms both baselines. FLUX.1-dev with reduced steps degrades image quality and exhibits over-saturation artifacts, while RALU introd… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparisons on latent-space image generation (fine-tuned). We compare our method against default-step generation, reduced-step native-resolution generation on Z-Image [5] matched to similar speedups. Our fine-tuned model (Ours∗ ) achieves even higher image quality compared to our training-free acceleration variant and outperforms the reduced-step baseline. 32 [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparisons on latent-space image generation (fine-tuned). We compare our method against default-step generation, reduced-step native-resolution generation on Z-Image [5] matched to similar speedups. Our fine-tuned model (Ours∗ ) achieves even higher image quality compared to our training-free acceleration variant and outperforms the reduced-step baseline. 33 [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparisons on latent-space image generation (fine-tuned). We compare our method against default-step generation, reduced-step native-resolution generation on Z-Image [5] matched to similar speedups. Our fine-tuned model (Ours∗ ) achieves even higher image quality compared to our training-free acceleration variant and outperforms the reduced-step baseline. 34 [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparisons on pixel-space image generation. We compare our method against default-step generation and reduced-step native-resolution generation on PixelGen [55], matched to comparable speedups. An asterisk (Ours∗ ) marks the fine-tuned model. Our method achieves similar quality to full-resolution generation while attaining higher speedups. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparisons on pixel-space image generation. We compare our method against default-step generation and reduced-step native-resolution generation on PixelGen [55], matched to comparable speedups. An asterisk (Ours∗ ) marks the fine-tuned model. Our method achieves similar quality to full-resolution generation while attaining higher speedups. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparisons on pixel-space image generation. We compare our method against default-step generation and reduced-step native-resolution generation on PixelGen [55], matched to comparable speedups. An asterisk (Ours∗ ) marks the fine-tuned model. Our method achieves similar quality to full-resolution generation while attaining higher speedups. 37 [PITH_FULL_IMAGE:figures/full_fig_p037_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative ablation on TΦ. We see that FFT leads to overly smooth results while DCT and DWT attain similar image quality. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative ablation on δ. We observe that increasing δ improves efficiency, but results in ghosting and halo artifacts near detailed edges. 42 [PITH_FULL_IMAGE:figures/full_fig_p042_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative ablation on S. We find that increasing S leads to marginal speedup improve￾ments and little image quality degradation. 43 [PITH_FULL_IMAGE:figures/full_fig_p043_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Texture editing results. Our frequency-based editing framework outperforms SDEdit, enabling high-fidelity texture transfer while preserving the geometric structure of the input image. 45 [PITH_FULL_IMAGE:figures/full_fig_p045_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Texture editing results. Our frequency-based editing framework outperforms SDEdit, enabling high-fidelity texture transfer while preserving the geometric structure of the input image. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Texture editing results. Our frequency-based editing framework outperforms SDEdit, enabling high-fidelity texture transfer while preserving the geometric structure of the input image. 47 [PITH_FULL_IMAGE:figures/full_fig_p047_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Texture editing results. Our frequency-based editing framework outperforms SDEdit, enabling high-fidelity texture transfer while preserving the geometric structure of the input image. 48 [PITH_FULL_IMAGE:figures/full_fig_p048_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Artistic stylization results. Aside from texture editing, our frequency-based editing approach also supports artistic stylization given stylistic descriptions and a representative artist. 49 [PITH_FULL_IMAGE:figures/full_fig_p049_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Effect of TΦ on image editing. FFT–based editing leads to overly-smooth and hazy results; DCT- and DWT-based editing achieve similar editing quality. 50 [PITH_FULL_IMAGE:figures/full_fig_p050_23.png] view at source ↗
read the original abstract

Diffusion models have been shown to implicitly generate visual content autoregressively in the frequency domain, where low-frequency components are generated earlier in the denoising process while high-frequency details emerge only in later timesteps. This structure offers a natural opportunity for efficient generation, as high-resolution computation on noise-dominated frequencies is largely redundant. We propose Spectral Progressive Diffusion, a general framework that progressively grows resolution along the denoising trajectory of pretrained diffusion models. To this end, we develop a spectral noise expansion mechanism and derive an optimal resolution schedule from the model's power spectrum. Our framework supports training-free acceleration and a novel fine-tuning recipe that further improves efficiency and quality. We demonstrate significant speedups on state-of-the-art pretrained image and video generation models while preserving visual quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Spectral Progressive Diffusion, a framework for accelerating pretrained diffusion models for image and video generation by progressively growing resolution along the denoising trajectory. It develops a spectral noise expansion mechanism and derives an optimal resolution schedule from the model's power spectrum, supporting both training-free acceleration and a fine-tuning recipe, with claims of significant speedups on SOTA models while preserving visual quality.

Significance. If the central claims hold, the work provides a general, practical approach to reducing the computational cost of high-resolution generation without retraining, which is valuable given the expense of diffusion-based image and video models. The emphasis on leveraging implicit frequency-domain structure in pretrained models and the dual training-free/fine-tuning support are strengths that could influence efficiency-focused extensions in generative modeling.

major comments (2)
  1. [§3] §3 (spectral noise expansion mechanism): The claim that the mechanism enables training-free application of the fixed pretrained denoiser requires that the expanded noise at each resolution transition exactly matches the marginal distribution (variance schedule and cross-frequency correlations) of the original high-resolution forward process at that timestep. The abstract and method description do not provide an explicit verification or derivation showing this preservation, raising a correctness risk for the subsequent denoising steps.
  2. [§4] §4 (optimal resolution schedule derivation): The schedule is stated to be derived from the model's power spectrum, but it is unclear whether the derivation operates under the exact forward-process marginals or relies on empirical averages; if the latter, the schedule may introduce model-specific fitting that undermines the generalizability of the training-free speedup claim.
minor comments (2)
  1. [Abstract] The abstract and introduction would benefit from explicit quantitative results (e.g., speedup factors, FID or perceptual metrics on specific models like Stable Diffusion or video variants) to ground the 'significant speedups' and 'preserving visual quality' claims.
  2. [Method] Notation for the power spectrum and resolution schedule should be defined with equations early in the method section to improve clarity for readers tracking the frequency-domain arguments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback on our manuscript. We have carefully addressed each major comment below with clarifications and planned revisions to improve the rigor and clarity of the presentation.

read point-by-point responses

  1. Referee: [§3] §3 (spectral noise expansion mechanism): The claim that the mechanism enables training-free application of the fixed pretrained denoiser requires that the expanded noise at each resolution transition exactly matches the marginal distribution (variance schedule and cross-frequency correlations) of the original high-resolution forward process at that timestep. The abstract and method description do not provide an explicit verification or derivation showing this preservation, raising a correctness risk for the subsequent denoising steps.

    Authors: We thank the referee for identifying this important aspect of the correctness argument. The current manuscript presents the spectral noise expansion mechanism but does not include a self-contained derivation of marginal preservation. In the revised manuscript we will add a formal derivation in §3 showing that the expansion, by construction via the Fourier basis and power-spectrum scaling, exactly reproduces the variance schedule and cross-frequency covariances of the high-resolution forward process at the transition timestep. We will also include a short verification experiment in the appendix that empirically confirms the distributional match before and after expansion. revision: yes

  2. Referee: [§4] §4 (optimal resolution schedule derivation): The schedule is stated to be derived from the model's power spectrum, but it is unclear whether the derivation operates under the exact forward-process marginals or relies on empirical averages; if the latter, the schedule may introduce model-specific fitting that undermines the generalizability of the training-free speedup claim.

    Authors: We appreciate the referee’s concern about the theoretical grounding and generalizability. The derivation in §4 starts from the exact forward-process marginals and uses the power spectrum to identify the timestep at which high-frequency energy falls below a noise-dominated threshold; the schedule is therefore analytic with respect to those marginals. In practice the power spectrum is estimated once from the pretrained model, but this estimation is not a learned fitting procedure and does not alter the underlying marginals. We will expand §4 to make this distinction explicit, add a short proof sketch linking the schedule directly to the marginal variance expressions, and include a brief discussion of why the same procedure applies to any diffusion model whose frequency-generation ordering is consistent with the observed power-spectrum decay. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper states it develops a spectral noise expansion mechanism and derives an optimal resolution schedule from the model's power spectrum. This uses the pretrained model's characteristics as an external input for the schedule rather than reducing the central result to a self-referential fit or self-citation by construction. No equations or steps in the abstract demonstrate that a 'prediction' equals its own fitted input or that a uniqueness claim collapses to prior author work. The framework is positioned as training-free acceleration on existing models, with the power spectrum providing independent frequency-domain structure. This is the common case of a self-contained method against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The approach rests on the domain observation that diffusion models generate low frequencies first. The spectral noise expansion is a newly introduced mechanism without independent evidence outside the paper. No explicit free parameters are named in the abstract.

axioms (1)
  • domain assumption Diffusion models implicitly generate visual content autoregressively in the frequency domain, with low-frequency components generated earlier.
    Directly stated in the abstract as the foundational observation enabling the method.
invented entities (1)
  • spectral noise expansion mechanism no independent evidence
    purpose: To enable progressive resolution growth along the denoising trajectory
    Newly proposed component of the framework; no external falsifiable evidence provided in abstract.

pith-pipeline@v0.9.0 · 5651 in / 1090 out tokens · 50067 ms · 2026-05-21T07:48:36.978396+00:00 · methodology

Review history (2 revisions) →

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

114 extracted references · 114 canonical work pages · 7 internal anchors

  1. [1]

    Ahmed, T

    N. Ahmed, T. Natarajan, and K. R. Rao. Discrete cosine transform.IEEE Transactions on Computers, C-23(1):90–93, 1974

    work page 1974

  2. [2]

    Longformer: The Long-Document Transformer

    Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer, 2020. URLhttps://arxiv.org/abs/2004.05150

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  3. [3]

    Spectral analysis of diffusion models with application to schedule design

    Roi Benita, Miki Elad, and Joseph Keshet. Spectral analysis of diffusion models with application to schedule design. InAdv. Neural Inform. Process. Syst., volume 38, pages 2073–2127, 2026

    work page 2073

  4. [4]

    Token merging for fast stable diffusion

    Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. InIEEE Conf. Comput. Vis. Pattern Recog., pages 4599–4603, 2023

    work page 2023

  5. [5]

    Z-Image: An efficient image generation foundation model with single-stream diffusion transformer, 2025

    Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven Hoi, et al. Z-Image: An efficient image generation foundation model with single-stream diffusion transformer, 2025. URL https://arxiv.org/abs/2511. 22699

    work page 2025

  6. [6]

    Spectral regularization for diffusion models, 2026

    Satish Chandran, Nicolas Roque dos Santos, Yunshu Wu, Greg Ver Steeg, and Evangelos Papalexakis. Spectral regularization for diffusion models, 2026. URLhttps://arxiv.org/abs/2603.02447

    work page arXiv 2026

  7. [7]

    δ-DiT: A training-free acceleration method tailored for diffusion transformers, 2024

    Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. δ-DiT: A training-free acceleration method tailored for diffusion transformers, 2024. URLhttps://arxiv.org/abs/2406.01125

    work page arXiv 2024

  8. [8]

    Rethinking attention with performers

    Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. InInt. Conf. Learn. Represent., 2021

    work page 2021

  9. [9]

    Diffusion is spectral autoregression

    Sander Dieleman. Diffusion is spectral autoregression. Blog post, September 2024. URL https: //sander.ai/2024/09/02/spectral-autoregression.html

    work page 2024

  10. [10]

    DemoFusion: Democratising high-resolution image generation with no $$$

    Ruoyi Du, Dongliang Chang, et al. DemoFusion: Democratising high-resolution image generation with no $$$. InIEEE Conf. Comput. Vis. Pattern Recog., 2024

    work page 2024

  11. [11]

    Flow along the k-amplitude for generative modeling, 2025

    Weitao Du, Shuning Chang, Jiasheng Tang, Yu Rong, Fan Wang, and Shengchao Liu. Flow along the k-amplitude for generative modeling, 2025. URLhttps://arxiv.org/abs/2504.19353

    work page arXiv 2025

  12. [12]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. InInt. Conf. Mach. Learn., 2024

    work page 2024

  13. [13]

    Spectrally-guided diffusion noise schedules

    Carlos Esteves and Ameesh Makadia. Spectrally-guided diffusion noise schedules. InInt. Conf. Mach. Learn., 2026

    work page 2026

  14. [14]

    A fourier space perspective on diffusion models, 2025

    Fabian Falck, Teodora Pandeva, Kiarash Zahirnia, Rachel Lawrence, Richard Turner, Edward Meeds, Javier Zazo, and Sushrut Karmalkar. A fourier space perspective on diffusion models, 2025. URL https://arxiv.org/abs/2505.11278

    work page arXiv 2025

  15. [15]

    Vchitect-2.0: Parallel transformer for scaling up video diffusion models,

    Weichen Fan, Chenyang Si, Junhao Song, Zhenyu Yang, Yinan He, Long Zhuo, Ziqi Huang, Ziyue Dong, Jingwen He, Dongwei Pan, et al. Vchitect-2.0: Parallel transformer for scaling up video diffusion models,

    work page

  16. [16]

    URLhttps://arxiv.org/abs/2501.08453

    work page arXiv

  17. [17]

    Attend to not attended: Structure-then-detail token merging for post-training dit acceleration

    Haipeng Fang, Sheng Tang, Juan Cao, Enshuo Zhang, Fan Tang, and Tong-Yee Lee. Attend to not attended: Structure-then-detail token merging for post-training dit acceleration. InIEEE Conf. Comput. Vis. Pattern Recog., pages 18083–18092, 2025

    work page 2025

  18. [18]

    Firmin Didot, 1822

    Jean-Baptiste Joseph Fourier.Théorie Analytique de la Chaleur. Firmin Didot, 1822

    work page

  19. [19]

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

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. GenEval: An object-focused framework for evaluating text-to-image alignment. InAdv. Neural Inform. Process. Syst., 2023

    work page 2023

  20. [20]

    Matryoshka diffusion models

    Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Joshua M Susskind, and Navdeep Jaitly. Matryoshka diffusion models. InInt. Conf. Learn. Represent., 2023

    work page 2023

  21. [21]

    Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation

    Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. InEur. Conf. Comput. Vis., 2024. 11

    work page 2024

  22. [22]

    Wavelet score-based generative modeling

    Florentin Guth, Simon Coste, Valentin De Bortoli, and Stephane Mallat. Wavelet score-based generative modeling. InAdv. Neural Inform. Process. Syst., volume 35, pages 478–491, 2022

    work page 2022

  23. [23]

    Zur theorie der orthogonalen funktionensysteme.Mathematische Annalen, 69(3):331–371, 1910

    Alfréd Haar. Zur theorie der orthogonalen funktionensysteme.Mathematische Annalen, 69(3):331–371, 1910

    work page 1910

  24. [24]

    Infinity: Scaling bitwise AutoRegressive modeling for high-resolution image synthesis

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise AutoRegressive modeling for high-resolution image synthesis. InIEEE Conf. Comput. Vis. Pattern Recog., 2025

    work page 2025

  25. [25]

    Agglomerative token clustering

    Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Agglomerative token clustering. InEur. Conf. Comput. Vis., pages 200–218, 2024

    work page 2024

  26. [26]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAdv. Neural Inform. Process. Syst., 2020

    work page 2020

  27. [27]

    Cascaded diffusion models for high fidelity image generation.J

    Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation.J. Mach. Learn. Res., 23(47):1–33, 2022

    work page 2022

  28. [28]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInt. Conf. Learn. Represent., 2022

    work page 2022

  29. [29]

    T2I-CompBench: A comprehensive benchmark for open-world compositional text-to-image generation

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2I-CompBench: A comprehensive benchmark for open-world compositional text-to-image generation. InAdv. Neural Inform. Process. Syst., 2023

    work page 2023

  30. [30]

    Wavedm: Wavelet-based diffusion models for image restoration.IEEE Trans

    Yi Huang, Jiancheng Huang, Jianzhuang Liu, Mingfu Yan, Yu Dong, Jiaxi Lv, Chaoqi Chen, and Shifeng Chen. Wavedm: Wavelet-based diffusion models for image restoration.IEEE Trans. Multimedia, 26: 7058–7073, 2024

    work page 2024

  31. [31]

    Spectralar: Spectral autoregressive visual generation

    Yuanhui Huang, Weiliang Chen, Wenzhao Zheng, Yueqi Duan, Jie Zhou, and Jiwen Lu. Spectralar: Spectral autoregressive visual generation. InInt. Conf. Comput. Vis., pages 15842–15852, 2025

    work page 2025

  32. [32]

    VBench: Comprehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. InIEEE Conf. Comput. Vis. Pattern Recog., 2024

    work page 2024

  33. [33]

    Training-free mixed-resolution latent upsampling for spatially accelerated diffusion transformers, 2026

    Wongi Jeong, Kyungryeol Lee, Hoigi Seo, and Se Young Chun. Training-free mixed-resolution latent upsampling for spatially accelerated diffusion transformers, 2026. URLhttps://arxiv.org/abs/2507. 08422

    work page 2026

  34. [34]

    Pyramidal flow matching for efficient video generative modeling

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. InInt. Conf. Learn. Represent., 2025

    work page 2025

  35. [35]

    Subspace diffusion generative models

    Bowen Jing, Gabriele Corso, Renato Berlinghieri, and Tommi Jaakkola. Subspace diffusion generative models. InEur. Conf. Comput. Vis., pages 274–289, 2022

    work page 2022

  36. [36]

    Kapfer, K

    C. Kapfer, K. Stine, B. Narasimhan, C. Mentzel, and E. Candès. Marlowe: Stanford’s GPU-based computational instrument, 2025. URLhttps://doi.org/10.5281/zenodo.14751899

    work page doi:10.5281/zenodo.14751899 2025

  37. [37]

    Transformers are rnns: Fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. InInt. Conf. Mach. Learn., pages 5156–5165. PMLR, 2020

    work page 2020

  38. [38]

    Token fusion: Bridging the gap between token pruning and token merging

    Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1383–1392, 2024

    work page 2024

  39. [39]

    DiffuseHigh: Training-free progressive high-resolution image synthesis through structure guidance

    Younghyun Kim, Geunmin Hwang, Junyu Zhang, and Eunbyung Park. DiffuseHigh: Training-free progressive high-resolution image synthesis through structure guidance. InAAAI, 2025

    work page 2025

  40. [40]

    Kingma, Tim Salimans, Ben Poole, and Jonathan Ho

    Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. InAdv. Neural Inform. Process. Syst., 2021

    work page 2021

  41. [41]

    Black Forest Labs. FLUX. Software repository, 2024. URL https://github.com/ black-forest-labs/flux. 12

    work page 2024

  42. [42]

    FLUX.2: Frontier Visual Intelligence

    Black Forest Labs. FLUX.2: Frontier Visual Intelligence. Blog post, 2025. URL https://bfl.ai/ blog/flux-2

    work page 2025

  43. [43]

    Beta Sampling is All You Need: Efficient Image Generation Strategy for Diffusion Models using Stepwise Spectral Analysis

    Haeil Lee, Hansang Lee, Seoyeon Gye, and Junmo Kim. Beta sampling is all you need: Efficient image generation strategy for diffusion models using stepwise spectral analysis, 2024. URL https: //arxiv.org/abs/2407.12173

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  44. [44]

    Local representative token guided merging for text-to-image generation, 2025

    Min-Jeong Lee, Hee-Dong Kim, and Seong-Whan Lee. Local representative token guided merging for text-to-image generation, 2025. URLhttps://arxiv.org/abs/2507.12771

    work page arXiv 2025

  45. [45]

    Frecas: Efficient higher-resolution image generation via frequency-aware cascaded sampling

    Ruihuang Li, Lei Zhang, et al. Frecas: Efficient higher-resolution image generation via frequency-aware cascaded sampling. InInt. Conf. Learn. Represent., volume 2025, pages 6400–6412, 2025

    work page 2025

  46. [46]

    Radial attention: O(nlogn) sparse attention with energy decay for long video generation

    Xingyang Li, Muyang Li, Tianle Cai, Haocheng Xi, Shuo Yang, Yujun Lin, Lvmin Zhang, Songlin Yang, Jinbo Hu, Kelly Peng, et al. Radial attention: O(nlogn) sparse attention with energy decay for long video generation. InAdv. Neural Inform. Process. Syst., 2025

    work page 2025

  47. [47]

    Wf-vae: Enhancing video vae by wavelet-driven energy flow for latent video diffusion model

    Zongjian Li, Bin Lin, Yang Ye, Liuhan Chen, Xinhua Cheng, Shenghai Yuan, and Li Yuan. Wf-vae: Enhancing video vae by wavelet-driven energy flow for latent video diffusion model. InIEEE Conf. Comput. Vis. Pattern Recog., pages 17778–17788, 2025

    work page 2025

  48. [48]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. InEur. Conf. Comput. Vis., 2014

    work page 2014

  49. [49]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInt. Conf. Learn. Represent., 2023

    work page 2023

  50. [50]

    Timestep embedding tells: It’s time to cache for video diffusion model, 2024

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It’s time to cache for video diffusion model, 2024. URL https://arxiv.org/abs/2411.19108

    work page arXiv 2024

  51. [51]

    FreqCa: Accelerating diffusion models via frequency-aware caching, 2025

    Jiacheng Liu, Peiliang Cai, Qinming Zhou, Yuqi Lin, Deyang Kong, Benhao Huang, Yupei Pan, Haowen Xu, Chang Zou, Junshu Tang, Shikang Zheng, and Linfeng Zhang. FreqCa: Accelerating diffusion models via frequency-aware caching, 2025. URLhttps://arxiv.org/abs/2510.08669

    work page arXiv 2025

  52. [52]

    From reusing to forecasting: Accelerating diffusion models with TaylorSeers

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, and Linfeng Zhang. From reusing to forecasting: Accelerating diffusion models with TaylorSeers. InInt. Conf. Comput. Vis., 2025

    work page 2025

  53. [53]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InInt. Conf. Learn. Represent., 2023

    work page 2023

  54. [54]

    ToMA: Token merge with attention for diffusion models

    Wenbo Lu, Shaoyi Zheng, Yuxuan Xia, and Shengjie Wang. ToMA: Token merge with attention for diffusion models. InInt. Conf. Mach. Learn., 2025

    work page 2025

  55. [55]

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference, 2023. URLhttps://arxiv.org/abs/2310.04378

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  56. [56]

    Deepcache: Accelerating diffusion models for free

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In IEEE Conf. Comput. Vis. Pattern Recog., 2024

    work page 2024

  57. [57]

    PixelGen: Improving Pixel Diffusion with Perceptual Supervision

    Zehong Ma, Ruihan Xu, and Shiliang Zhang. PixelGen: Pixel diffusion beats latent diffusion with perceptual loss, 2026. URLhttps://arxiv.org/abs/2602.02493

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  58. [58]

    SDEdit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. InInt. Conf. Learn. Represent., 2022

    work page 2022

  59. [59]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a “completely blind” image quality analyzer.IEEE Sign. Process. Letters, 20(3):209–212, 2013. doi: 10.1109/LSP.2012.2227726

    work page doi:10.1109/lsp.2012.2227726 2013

  60. [60]

    Scale space diffusion, 2026

    Soumik Mukhopadhyay, Prateksha Udhayanan, and Abhinav Shrivastava. Scale space diffusion, 2026. URLhttps://arxiv.org/abs/2603.08709

    work page arXiv 2026

  61. [61]

    DCTdiff: Intriguing properties of image generative modeling in the DCT space

    Mang Ning, Mingxiao Li, Jianlin Su, Jia Haozhe, Lanmiao Liu, Martin Benes, Wenshuo Chen, Albert Ali Salah, and Itir Onal Ertugrul. DCTdiff: Intriguing properties of image generative modeling in the DCT space. InInt. Conf. Mach. Learn., volume 267, pages 46498–46524. PMLR, 2025. 13

    work page 2025

  62. [62]

    NVIDIA, Yuval Atzmon, Maciej Bala, Yogesh Balaji, Tiffany Cai, Yin Cui, Jiaojiao Fan, Yunhao Ge, Siddharth Gururani, Jacob Huffman, Ronald Isaac, Pooya Jannaty, Tero Karras, Grace Lam, J. P. Lewis, Aaron Licata, Yen-Chen Lin, Ming-Yu Liu, Qianli Ma, Arun Mallya, Ashlee Martino-Tarr, Doug Mendez, Seungjun Nah, Chris Pruett, Fitsum Reda, Jiaming Song, Ting-...

    work page arXiv 2024

  63. [63]

    Certain topics in telegraph transmission theory.Transactions of the American Institute of Electrical Engineers, 47(2):617–644, 1928

    Harry Nyquist. Certain topics in telegraph transmission theory.Transactions of the American Institute of Electrical Engineers, 47(2):617–644, 1928

    work page 1928

  64. [64]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InInt. Conf. Comput. Vis., 2023

    work page 2023

  65. [65]

    Wavelet diffusion models are fast and scalable image generators

    Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In IEEE Conf. Comput. Vis. Pattern Recog., 2023

    work page 2023

  66. [66]

    DiMSUM: Diffusion Mamba – a scalable and unified spatial-frequency method for image generation

    Hao Phung, Quan Dao, Trung Dao, Hoang Phan, Dimitris Metaxas, and Anh Tran. DiMSUM: Diffusion Mamba – a scalable and unified spatial-frequency method for image generation. InAdv. Neural Inform. Process. Syst., 2024

    work page 2024

  67. [67]

    FreeScale: Unleashing the resolution of diffusion models via tuning-free scale fusion

    Haonan Qiu et al. FreeScale: Unleashing the resolution of diffusion models via tuning-free scale fusion. In Int. Conf. Comput. Vis., 2025

    work page 2025

  68. [68]

    FlowAR: Scale-wise autoregressive image generation meets flow matching

    Sucheng Ren, Qihang Yu, Ju He, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. FlowAR: Scale-wise autoregressive image generation meets flow matching. InInt. Conf. Mach. Learn., 2025

    work page 2025

  69. [69]

    Generative modelling with inverse heat dissipation

    Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. InInt. Conf. Learn. Represent., 2023

    work page 2023

  70. [70]

    Origins of scaling in natural images.Vis

    Daniel L Ruderman. Origins of scaling in natural images.Vis. Res., 37(23):3385–3398, 1997

    work page 1997

  71. [71]

    Pyramidal denoising diffusion probabilistic models, 2022

    Dohoon Ryu and Jong Chul Ye. Pyramidal denoising diffusion probabilistic models, 2022. URL https: //arxiv.org/abs/2208.01864

    work page arXiv 2022

  72. [72]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. InInt. Conf. Learn. Represent., 2022

    work page 2022

  73. [73]

    Claude E. Shannon. Communication in the presence of noise.Proceedings of the IRE, 37(1):10–21, 1949

    work page 1949

  74. [74]

    Latent Wavelet Diffusion For Ultra-High-Resolution Image Synthesis

    Luigi Sigillo, Shengfeng He, and Danilo Comminiello. Latent wavelet diffusion for ultra-high-resolution image synthesis, 2025. URLhttps://arxiv.org/abs/2506.00433

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  75. [75]

    Improving the diffusability of autoencoders

    Ivan Skorokhodov, Sharath Girish, Benran Hu, Willi Menapace, Yanyu Li, Rameen Abdal, Sergey Tulyakov, and Aliaksandr Siarohin. Improving the diffusability of autoencoders. InInt. Conf. Mach. Learn., 2025

    work page 2025

  76. [76]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. InInt. Conf. Learn. Represent., 2021

    work page 2021

  77. [77]

    Lssgen: Leveraging latent space scaling in flow and diffusion for efficient text to image generation

    Jyun-Ze Tang, Chih-Fan Hsu, Jeng-Lin Li, Ming-Ching Chang, and Wei-Chao Chen. Lssgen: Leveraging latent space scaling in flow and diffusion for efficient text to image generation. InInt. Conf. Comput. Vis., pages 5048–5057, 2025

    work page 2025

  78. [79]

    Relay diffusion: Unifying diffusion process across resolutions for image synthesis

    Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. InInt. Conf. Learn. Represent.,

    work page

  79. [80]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. InAdv. Neural Inform. Process. Syst., 2024

    work page 2024

  80. [81]

    Training-free diffusion acceleration with bot- tleneck sampling.arXiv preprint arXiv:2503.18940,

    Ye Tian, Xin Xia, Yuxi Ren, Shanchuan Lin, Xing Wang, Xuefeng Xiao, Yunhai Tong, Ling Yang, and Bin Cui. Training-free diffusion acceleration with bottleneck sampling, 2025. URL https://arxiv.org/ abs/2503.18940

    work page arXiv 2025

Showing first 80 references.