Pith. sign in

REVIEW 3 major objections 4 minor 131 references

This survey establishes pixel-space diffusion transformers as a distinct, scalable class of end-to-end generative models that removes the fixed VAE bottleneck by denoising directly in raw pixel space.

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 · deepseek-v4-flash

2026-08-01 17:32 UTC pith:54UJ2QT7

load-bearing objection A useful survey with a solid taxonomy, but the 'first systematic' claim is unaudited and needs a methodology section. the 3 major comments →

arxiv 2607.17585 v2 pith:54UJ2QT7 submitted 2026-07-20 cs.CV

Pixel-Space Diffusion Transformers

Renye Yan , Jikang Cheng , You Wu , Ling Liang , Wei Peng , Athanasios V. Vasilakos , Qingyu Zhao , Yu Zhang
show 3 more authors
Yimao Cai Kilian M. Pohl Guoying Zhao
This is my paper
classification cs.CV
keywords pixel-space diffusiondiffusion transformerslatent diffusion modelsgenerative modelingunified multimodal modelsimage generationflow matchingsurvey
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.

The paper argues that pixel-space diffusion transformers (pDiTs) form a distinct, scalable class of generative models that removes the fixed VAE bottleneck of latent diffusion by defining the denoising trajectory directly on raw pixels. It positions this shift as resolving a learnability-fidelity conflict through modern training recipes, direct clean-image prediction, and structured Transformer architectures. As the first survey dedicated to pDiTs, it organizes the field into seven architectural categories across three research dimensions — architecture, continuous generative mechanisms, and unified multimodal modeling — and identifies open challenges including quadratic attention cost, granularity conflicts, and multi-task gradient interference. A sympathetic reader would care because the paper maps how image generation could move toward end-to-end, VAE-free vision foundation models.

Core claim

The paper's central claim is that pDiTs are not merely a return to early pixel-space diffusion but an independent, scalable paradigm: 'pixel space' means the diffusion state, prediction target, and supervision are defined in the raw image domain, while patchification is only a computational tokenization strategy, not a frozen visual compressor. It asserts that this removes an irreversible information bottleneck, allows single-stage end-to-end optimization, and provides a shared token space in which text, pixels, and task conditions can be jointly modeled by one Transformer. The survey substantiates this through a seven-category architectural taxonomy, an analysis of DDPM and flow-matching fo

What carries the argument

The load-bearing object is the pDiT formulation: a noisy image is patchified, each patch is mapped by a learnable embedding into tokens, a Transformer backbone models global dependencies under time and text conditions, and an end-to-end decoding head predicts the clean image in pixel space. The survey's analytic grid is a seven-category taxonomy — single-stream large-patch, hierarchical/hourglass, global-local decoupled, frequency-decoupled, implicit neural field decoding, cross-scale semantic anchoring, and shared-token unified multimodal architectures — with flow matching and DDPM as the continuous trajectory mathematics and the O(N^2 d) attention cost as the central scaling constraint.

Load-bearing premise

The survey's usefulness rests on the assumption that its selection of representative methods and its seven-category taxonomy faithfully cover the pDiT landscape; if important methods are omitted or miscategorized, the systematic and first-survey claims lose force.

What would settle it

A concrete check: compile an independent list of pixel-space diffusion transformer papers with stated noise schedules and loss weightings, then test whether each fits one of the seven Table I categories; any substantial class that does not fit, or a matched-compute benchmark where latent diffusion beats pixel-space models specifically on text rendering and edge fidelity, would undercut the central taxonomy and the claim that the VAE bottleneck is removed.

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

If this is right

  • Fixed VAE or vision-foundation tokenizers become an optional interface rather than a quality ceiling, since raw-pixel supervision can preserve textures, edges, and text that compression discards.
  • End-to-end training can jointly optimize representation and generation, reducing the reconstruction-generation mismatch inherent in two-stage latent diffusion.
  • Unified multimodal models can be built on a shared token space where pixels, text, and task conditions are processed by a single Transformer.
  • Structured computation — large patches, hierarchy, global-local decoupling, frequency separation, and implicit decoding — makes high-resolution pixel-space generation tractable.
  • Evaluation must move beyond a single FID score or latency figure to include text fidelity, structural preservation, edit locality, and compute-fidelity Pareto comparisons.

Where Pith is reading between the lines

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

  • If the pDiT advantages hold, the most promising architectures will likely combine several taxonomy mechanisms — for example, global-local decoupling with frequency-aware scheduling and dynamic token granularity — rather than rely on a single design.
  • The learnability-fidelity conflict suggests a testable scaling prediction: at sufficient compute and with clean-image prediction or structured supervision, pixel-space models should match or surpass latent models on fine-detail metrics under matched training FLOPs.
  • The observed risk that token-wise representation alignment degrades diversity in pixel space implies that future guidance will shift toward relation-level or stage-adaptive constraints; a direct comparison of Gram-matrix alignment versus token-wise alignment on diversity and fidelity would test this.
  • If the shared-token unified paradigm is realized, multi-task gradient interference would need active management; monitoring gradient cosine-similarity conflicts during training of such a unified model could predict when understanding and generation objectives fight each other.

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 / 4 minor

Summary. The paper presents a survey of Pixel-Space Diffusion Transformers (pDiTs), arguing that this emerging class of generative models—defined by diffusion and supervision in the raw image space, with Transformer backbones and no fixed VAE/VQ tokenizer—offers a distinct and scalable alternative to latent diffusion. The survey covers theoretical foundations (DDPM, flow matching, complexity analysis), a seven-category architectural taxonomy (Table I), unified multimodal modeling, applications (image, video, 3D, medical), discussion, and future challenges. The central claim is that this is 'the first survey dedicated to the systematic review and comprehensive analysis of pixel-space diffusion models' (Sec. I).

Significance. If the taxonomy and coverage are reliable, the paper would serve as a useful roadmap for a rapidly evolving area. The background mathematics (Eqs. 1–11) is standard and correctly reproduced, and the paper clearly organizes the field along three dimensions: architecture, continuous generative mechanisms, and unified multimodal modeling. It also identifies substantive open problems, such as the computation–fidelity trade-off, non-invasive semantic guidance, forgetting in unified models, and native pixel-space trajectory design. The survey is balanced in acknowledging the continued strengths of latent diffusion. However, the value of the survey as a 'systematic review' depends on the completeness and fidelity of its method selection and categorization, which are not auditable in the current manuscript.

major comments (3)
  1. [Sec. I and Sec. III-B (Table I)] The paper's central claim of being 'the first survey dedicated to the systematic review and comprehensive analysis of pixel-space diffusion models' is not supported by any stated methodology. The manuscript does not provide a search protocol, inclusion/exclusion criteria, database sources, or a time window, so the 'systematic' claim is unverifiable. Table I lists only one or two representative methods per category, but there is no justification for why these methods were chosen or why the seven categories are mutually exclusive and collectively exhaustive. Please add a methodology subsection describing the literature search and selection process, or soften the 'systematic/first' claim accordingly.
  2. [Sec. III and Sec. V-A] There is an internal inconsistency in the definition of pDiT. Section III defines pDiT as a Transformer-based architecture: 'the Transformer uses tokenized representations and attention mechanisms as its primary computational backbone.' However, Section V-A states that 'The earliest applications of pDiT' include Simple Diffusion [67] and SiD2 [57], both of which are U-Net-based pixel-space diffusion models, not Transformers. The same issue appears in Fig. 2 and the timeline. This conflates 'pixel-space diffusion' with 'pixel-space diffusion Transformers.' Please either restrict the application discussion to Transformer-based models or explicitly distinguish between pixel-space diffusion in general and the pDiT subset, and adjust the taxonomy and title accordingly.
  3. [Sec. V-A, Sec. V-B] The application sections mix pDiT-specific claims with general pixel-space diffusion results without sufficient discrimination. For example, BlazeEdit [123] is described as 'still following the latent diffusion paradigm rather than pDiT,' which is appropriately flagged, but other methods are listed as pDiT without checking whether they use a Transformer backbone. The survey would be strengthened by a consistent criterion for which methods enter the pDiT taxonomy and which are included only as adjacent pixel-space approaches. This is necessary to make the seven-category taxonomy in Table I a faithful representation of the pDiT landscape.
minor comments (4)
  1. [Fig. 2] Typo: 'PixWAorld' should be 'PixWorld'.
  2. [References] Reference [5] has an incomplete title: 'in computer vision' appears where a topic phrase is expected.
  3. [Eq. (33)] The equation formatting has stray line-break characters (\r\r) inside the alignment loss expression; please clean up the LaTeX.
  4. [Sec. II-B] The figure overview (Fig. 1) mentions 'Time Parameterization' as a background topic, but the background section does not discuss time parameterizations explicitly beyond DDPM/flow matching. Consider adding a short paragraph to align the overview with the content.

Circularity Check

0 steps flagged

No significant circularity: the survey's taxonomy and claims are literature-based and do not reduce to the paper's own inputs.

full rationale

This paper is a survey, not a derivation. The central claims are (i) that pixel-space diffusion transformers are a distinct scalable class of generative models and (ii) that this is the first systematic survey of that area. Neither claim is derived from an equation or fitted parameter within the paper. The theoretical background (DDPM, flow matching, attention complexity) consists of standard, externally established results presented for context, not used to predict a quantity that was already assumed. The Table I taxonomy and the three organizing dimensions are interpretative classifications of external methods; they are not presented as predictions and no method is scored or fitted against the taxonomy to produce a result. The self-citations (refs. [1], [10], [26], [37], [122]) appear only as general contextual references in the introduction and the privacy discussion; none of them is invoked to justify the taxonomy, the 'first survey' novelty claim, or any technical conclusion. The absence of stated inclusion/exclusion criteria for the surveyed methods is a legitimate audit concern about completeness and representativeness, but it is not a circularity: the survey's organizational structure is not equivalent to its own literature sample by construction. No step in the paper reduces, by definition or by self-citation, to its own inputs. Accordingly, the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The survey introduces no fitted parameters or new entities; its claim burden is carried by the accuracy and representativeness of its cited literature, and by standard diffusion/flow-matching math.

axioms (3)
  • domain assumption Cited primary works (Simple Diffusion, JiT, PixelDiT, HDiT, HiDream-O1-Image, etc.) are correctly described and their results are real.
    The survey's taxonomy and comparisons stand or fall on accurate representation of external papers; no independent verification is provided.
  • standard math The standard DDPM and flow-matching formulations (Eqs. 1-11) are standard mathematics as used in the literature.
    Background equations are standard; no new derivation is offered.
  • domain assumption The absence of an earlier dedicated survey of pixel-space diffusion models (the 'first survey' claim) is true.
    Not verifiable from the paper; depends on completeness of the authors' literature search.

pith-pipeline@v1.3.0-alltime-deepseek · 37911 in / 9790 out tokens · 93029 ms · 2026-08-01T17:32:42.989117+00:00 · methodology

0 comments
read the original abstract

Latent diffusion models (LDMs) enable efficient high-resolution image synthesis by denoising in a VAE-compressed latent space. However, fixed visual tokenizers can discard fine textures and structural details, while separate representation and diffusion training creates a mismatch between reconstruction and generation objectives. These limitations have renewed interest in pixel-space diffusion, which models raw pixels directly, removes the VAE bottleneck, and supports end-to-end optimization. This formulation better matches the demands of high-fidelity generation but introduces challenges in high-dimensional modeling, including noise scheduling, loss weighting, token efficiency, and scalable architecture design. Pixel-space modeling also offers a promising basis for unified multimodal systems: raw pixels, text, and task conditions can be represented in a shared token space and jointly processed by a single Transformer, narrowing the gap between visual understanding and generation. This paper reviews Pixel-Space Diffusion Transformers (pDiTs) from the perspectives of model architecture, continuous generative mechanisms, and unified multimodal modeling. We summarize representative methods, identify key technical challenges, and discuss future directions toward high-fidelity, end-to-end vision foundation models that integrate generation and understanding.

Figures

Figures reproduced from arXiv: 2607.17585 by Athanasios V. Vasilakos, Guoying Zhao, Jikang Cheng, Kilian M. Pohl, Ling Liang, Qingyu Zhao, Renye Yan, Wei Peng, Yimao Cai, You Wu, Yu Zhang.

Figure 1
Figure 1. Figure 1: Overview of the Survey Structure and Research Roadmap. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Timeline of the development of pixel diffusion models. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Taxonomy and Representative Methods of Latent Diffu [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of the Overall Pipelines and Key Differences between Latent-Space and Pixel-Space Diffusion Models. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

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

131 extracted references · 36 linked inside Pith

  1. [1]

    A sanity check for multi-in-domain face forgery detection in the real world,

    J. Cheng, R. Yan, Z. Yan, Y . Gan, X. Zhang, Z. Wang, W. Peng, and L. Liang, “A sanity check for multi-in-domain face forgery detection in the real world,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 21 306–21 315

    2026

  2. [2]

    Edge deep learning in computer vision and medical diagnostics: a com- prehensive survey,

    Y . Xu, T. M. Khan, Y . Song, and E. Meijering, “Edge deep learning in computer vision and medical diagnostics: a com- prehensive survey,”arXiv preprint arXiv:2605.06714, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    A review of pseudo-labeling for computer vision,

    P. Kage, J. Rothenberger, P. Andreadis, and D. Diochnos, “A review of pseudo-labeling for computer vision,”Journal of Artificial Intelligence Research, vol. 85, 2026

    2026

  4. [4]

    Intelligent recognition of emergency vehicles in congested traffic using computer vision,

    K. P. M. Ballesteros, C. M. L. D. Cruz, J. D. R. Magbanua, M. J. M. Mancenido, and L. V . Comia, “Intelligent recognition of emergency vehicles in congested traffic using computer vision,” in2026 6th International Conference on Image Pro- cessing and Capsule Networks (ICIPCN). IEEE, 2026, pp. 34–40

    2026

  5. [5]

    in computer vision,

    Y . Ji, W. Wu, H. Chen, and Z. Liu, “in computer vision,” Artificial Intelligence in Digital Image Processing: Theories, Methods, and Applications, p. 79, 2026

    2026

  6. [6]

    Attention mechanisms in computer vision: A survey,

    M.-H. Guo, T.-X. Xu, J.-J. Liu, Z.-N. Liu, P.-T. Jiang, T.- J. Mu, S.-H. Zhang, R. R. Martin, M.-M. Cheng, and S.-M. Hu, “Attention mechanisms in computer vision: A survey,” Computational visual media, vol. 8, no. 3, pp. 331–368, 2022

    2022

  7. [7]

    Generative adversarial networks in computer vision: A survey and taxonomy,

    Z. Wang, Q. She, and T. E. Ward, “Generative adversarial networks in computer vision: A survey and taxonomy,”ACM Computing Surveys (CSUR), vol. 54, no. 2, pp. 1–38, 2021

    2021

  8. [8]

    Picture perfect: Engaging customers with visual generative ai,

    M. Heitmann, T. P. Jansen, M. Reisenbichler, and D. A. Schweidel, “Picture perfect: Engaging customers with visual generative ai,”Journal of Marketing, vol. 90, no. 4, pp. 74–96, 2026

    2026

  9. [9]

    Around the world in 80 timesteps: A generative approach to global visual geolocation,

    N. Dufour, V . Kalogeiton, D. Picard, and L. Landrieu, “Around the world in 80 timesteps: A generative approach to global visual geolocation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 23 016–23 026

    2025

  10. [10]

    Entropy-adaptive diffusion policy optimization with dynamic step alignment,

    R. Yan, J. Cheng, Y . Gan, S. Sun, Y . Wu, Y . Yang, L. Ling, J. Lin, Y . Zhu, J. Zhouet al., “Entropy-adaptive diffusion policy optimization with dynamic step alignment,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 1924–1934

    2025

  11. [11]

    Deep generative modelling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive mod- els,

    S. Bond-Taylor, A. Leach, Y . Long, and C. G. Willcocks, “Deep generative modelling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive mod- els,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 11, pp. 7327–7347, 2021

    2021

  12. [12]

    Energy-based generative adversarial network,

    J. Zhao, M. Mathieu, and Y . LeCun, “Energy-based generative adversarial network,”arXiv preprint arXiv:1609.03126, 2016

    Pith/arXiv arXiv 2016

  13. [13]

    Nice: Non- linear independent components estimation,

    L. Dinh, D. Krueger, and Y . Bengio, “Nice: Non- linear independent components estimation,”arXiv preprint arXiv:1410.8516, 2014

    Pith/arXiv arXiv 2014

  14. [14]

    Wavenet: A generative model for raw audio,

    A. Van Den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, K. Kavukcuogluet al., “Wavenet: A generative model for raw audio,”arXiv preprint arXiv:1609.03499, vol. 12, no. 1, 2016

    Pith/arXiv arXiv 2016

  15. [15]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inProceedings of the IEEE international confer- ence on computer vision, 2017, pp. 2223–2232

    2017

  16. [16]

    Diffusion models in vision: A survey,

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah, “Diffusion models in vision: A survey,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 9, pp. 10 850–10 869, 2023

    2023

  17. [17]

    Mixture of global and local experts with diffusion transformer for controllable face generation,

    X. Zou, S. Zhang, X. Fu, Y . Li, K. Li, Y . Cao, C. Lang, P. Tao, and J. Xing, “Mixture of global and local experts with diffusion transformer for controllable face generation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  18. [18]

    Remasking 25 discrete diffusion models with inference-time scaling,

    G. Wang, Y . Schiff, S. Sahoo, and V . Kuleshov, “Remasking 25 discrete diffusion models with inference-time scaling,”Ad- vances in Neural Information Processing Systems, vol. 38, pp. 147 282–147 339, 2026

    2026

  19. [19]

    Why diffu- sion models don’t memorize: The role of implicit dynamical regularization in training,

    T. Bonnaire, R. Urfin, G. Biroli, and M. M ´ezard, “Why diffu- sion models don’t memorize: The role of implicit dynamical regularization in training,”Advances in Neural Information Processing Systems, vol. 38, pp. 141 266–141 286, 2026

    2026

  20. [20]

    Guiding a diffusion model with a bad version of itself,

    T. Karras, M. Aittala, T. Kynk ¨a¨anniemi, J. Lehtinen, T. Aila, and S. Laine, “Guiding a diffusion model with a bad version of itself,”Advances in Neural Information Processing Systems, vol. 37, pp. 52 996–53 021, 2024

    2024

  21. [21]

    Human motion diffusion model,

    G. Tevet, S. Raab, B. Gordon, Y . Shafir, D. Cohen-Or, and A. H. Bermano, “Human motion diffusion model,”arXiv preprint arXiv:2209.14916, 2022

    Pith/arXiv arXiv 2022

  22. [22]

    Diffusion models and representation learning: A survey,

    M. Fuest, P. Ma, M. Gui, J. Schusterbauer, V . T. Hu, and B. Ommer, “Diffusion models and representation learning: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  23. [23]

    Holod- iffusion: Training a 3d diffusion model using 2d images,

    A. Karnewar, A. Vedaldi, D. Novotny, and N. J. Mitra, “Holod- iffusion: Training a 3d diffusion model using 2d images,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 18 423–18 433

    2023

  24. [24]

    Klass: Kl-guided fast inference in masked diffusion models,

    S. H. Kim, S. Hong, H. Jung, Y . Park, and S.-Y . Yun, “Klass: Kl-guided fast inference in masked diffusion models,” Advances in Neural Information Processing Systems, vol. 38, pp. 92 267–92 301, 2026

    2026

  25. [25]

    Training-free constrained generation with stable diffusion models,

    S. Zampini, J. K. Christopher, L. Oneto, D. Anguita, and F. Fioretto, “Training-free constrained generation with stable diffusion models,”Advances in Neural Information Processing Systems, vol. 38, pp. 27 285–27 316, 2026

    2026

  26. [26]

    Do less, achieve more: Do we need every-step optimization for rl fine-tuning of diffusion models?

    R. Yan, J. Cheng, S. Sun, Y . Sun, Y . Wu, W. Peng, Z. Wang, L. Liang, J. Xing, and Y . Cai, “Do less, achieve more: Do we need every-step optimization for rl fine-tuning of diffusion models?” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 16 561– 16 571

    2026

  27. [27]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Om- mer, “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684– 10 695

    2022

  28. [28]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis,

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M ¨uller, J. Penna, and R. Rombach, “Sdxl: Improving latent diffusion models for high-resolution image synthesis,” inInter- national Conference on Learning Representations, vol. 2024, 2024, pp. 1862–1874

    2024

  29. [29]

    Align your latents: High-resolution video synthesis with latent diffusion models,

    A. Blattmann, R. Rombach, H. Ling, T. Dockhorn, S. W. Kim, S. Fidler, and K. Kreis, “Align your latents: High-resolution video synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 563–22 575

    2023

  30. [30]

    Stable video diffusion: Scaling latent video diffusion models to large datasets,

    A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kil- ian, D. Lorenz, Y . Levi, Z. English, V . V oleti, A. Lettset al., “Stable video diffusion: Scaling latent video diffusion models to large datasets,”arXiv preprint arXiv:2311.15127, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Diffusion models already have a semantic latent space,

    M. Kwon, J. Jeong, and Y . Uh, “Diffusion models already have a semantic latent space,”arXiv preprint arXiv:2210.10960, 2022

    Pith/arXiv arXiv 2022

  32. [32]

    Latentpaint: Image inpainting in latent space with diffusion models,

    C. Corneanu, R. Gadde, and A. M. Martinez, “Latentpaint: Image inpainting in latent space with diffusion models,” in Proceedings of the IEEE/CVF winter conference on applica- tions of computer vision, 2024, pp. 4334–4343

    2024

  33. [33]

    Unconditional latent diffusion models memorize patient imaging data,

    S. U. H. Dar, M. Seyfarth, I. Ayx, T. Papavassiliu, S. O. Schoenberg, R. M. Siepmann, F. C. Laqua, J. Kahmann, N. Frey, B. Baeßleret al., “Unconditional latent diffusion models memorize patient imaging data,”Nature biomedical engineering, vol. 10, no. 3, pp. 458–472, 2026

    2026

  34. [34]

    Ladder variational autoencoders,

    C. K. Sønderby, T. Raiko, L. Maaløe, S. K. Sønderby, and O. Winther, “Ladder variational autoencoders,”Advances in neural information processing systems, vol. 29, 2016

    2016

  35. [35]

    Longitudinal variational autoencoder,

    S. Ramchandran, G. Tikhonov, K. Kujanp ¨a¨a, M. Koskinen, and H. L¨ahdesm¨aki, “Longitudinal variational autoencoder,” inIn- ternational Conference on Artificial Intelligence and Statistics. PMLR, 2021, pp. 3898–3906

    2021

  36. [36]

    Gram- mar variational autoencoder,

    M. J. Kusner, B. Paige, and J. M. Hern ´andez-Lobato, “Gram- mar variational autoencoder,” inInternational conference on machine learning. PMLR, 2017, pp. 1945–1954

    2017

  37. [37]

    The exploration-exploitation dilemma revisited: An entropy perspective,

    R. Yan, Y . Gan, Y . Wu, L. Liang, J. Xing, Y . Cai, and R. Huang, “The exploration-exploitation dilemma revisited: An entropy perspective,”arXiv preprint arXiv:2408.09974, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    Mvae: Multimodal variational autoencoder for fake news detection,

    D. Khattar, J. S. Goud, M. Gupta, and V . Varma, “Mvae: Multimodal variational autoencoder for fake news detection,” inThe world wide web conference, 2019, pp. 2915–2921

    2019

  39. [39]

    Poisson variational autoencoder,

    H. Vafaii, D. Galor, and J. L. Yates, “Poisson variational autoencoder,”Advances in Neural Information Processing Sys- tems, vol. 37, pp. 44 871–44 906, 2024

    2024

  40. [40]

    Xidintfl-vae: Xgboost-based intrusion detection of imbalance network traffic via class-wise focal loss variational autoencoder,

    O. H. Abdulganiyu, T. A. Tchakoucht, Y . K. Saheed, and H. A. Ahmed, “Xidintfl-vae: Xgboost-based intrusion detection of imbalance network traffic via class-wise focal loss variational autoencoder,”The Journal of Supercomputing, vol. 81, no. 1, p. 16, 2025

    2025

  41. [41]

    Remaining useful life prediction based on interpretable serialized variational autoencoder: A drift- diffusion stochastic equation perspective,

    J. Zhang, K. Chen, R. He, T. Huang, J. Tian, S. Wu, P. Yan, and Y . Cheng, “Remaining useful life prediction based on interpretable serialized variational autoencoder: A drift- diffusion stochastic equation perspective,”IEEE Transactions on Industrial Informatics, 2026

    2026

  42. [42]

    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, 2023, pp. 4195–4205

    2023

  43. [43]

    All are worth words: A vit backbone for diffusion models,

    F. Bao, S. Nie, K. Xue, Y . Cao, C. Li, H. Su, and J. Zhu, “All are worth words: A vit backbone for diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 669–22 679

    2023

  44. [44]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers,

    N. Ma, M. Goldstein, M. S. Albergo, N. M. Boffi, E. Vanden- Eijnden, and S. Xie, “Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers,” in European Conference on Computer Vision. Springer, 2024, pp. 23–40

    2024

  45. [45]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M ¨uller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boeselet al., “Scaling rectified flow transformers for high-resolution image synthesis,” inForty-first international conference on machine learning, 2024

    2024

  46. [46]

    Exploring diffusion transformer designs via grafting,

    K. Chandrasegaran, M. Poli, D. Fu, D. Kim, L. M. Hadzic, M. Li, A. Gupta, S. Massaroli, A. Mirhoseini, J. C. Niebles et al., “Exploring diffusion transformer designs via grafting,” Advances in Neural Information Processing Systems, vol. 38, pp. 17 816–17 847, 2026

    2026

  47. [47]

    Scalable high-resolution pixel- space image synthesis with hourglass diffusion transformers,

    K. Crowson, S. A. Baumann, A. Birch, T. M. Abraham, D. Z. Kaplan, and E. Shippole, “Scalable high-resolution pixel- space image synthesis with hourglass diffusion transformers,” inForty-first International Conference on Machine Learning, 2024

    2024

  48. [48]

    Edify image: High- quality image generation with pixel space laplacian diffusion models,

    Y . Atzmon, M. Bala, Y . Balaji, T. Cai, Y . Cui, J. Fan, Y . Ge, S. Gururani, J. Huffman, R. Isaacet al., “Edify image: High- quality image generation with pixel space laplacian diffusion models,”arXiv preprint arXiv:2411.07126, 2024

    Pith/arXiv arXiv 2024

  49. [49]

    Pixelgen: Pixel diffusion beats latent diffusion with perceptual loss,

    Z. Ma, R. Xu, and S. Zhang, “Pixelgen: Pixel diffusion beats latent diffusion with perceptual loss,”arXiv preprint arXiv:2602.02493, 2026

    Pith/arXiv arXiv 2026

  50. [50]

    Pixnerd: Pixel neural field diffusion,

    S. Wang, Z. Gao, C. Zhu, W. Huang, and L. Wang, “Pixnerd: Pixel neural field diffusion,”arXiv preprint arXiv:2507.23268, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    Pixeldit: Pixel diffusion transformers for image generation,

    Y . Yu, W. Xiong, W. Nie, Y . Sheng, S. Liu, and J. Luo, “Pixeldit: Pixel diffusion transformers for image generation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 14 273–14 282

    2026

  52. [52]

    Deco: 26 Frequency-decoupled pixel diffusion for end-to-end image generation,

    Z. Ma, L. Wei, S. Wang, S. Zhang, and Q. Tian, “Deco: 26 Frequency-decoupled pixel diffusion for end-to-end image generation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 43 600– 43 610

    2026

  53. [53]

    Latent forcing: Reordering the diffusion trajectory for pixel-space image generation,

    A. Baade, E. R. Chan, K. Sargent, C. Chen, J. Johnson, E. Adeli, and L. Fei-Fei, “Latent forcing: Reordering the diffusion trajectory for pixel-space image generation,”arXiv preprint arXiv:2602.11401, 2026

    arXiv 2026

  54. [54]

    Your latent mask is wrong: Pixel-equivalent latent compositing for diffusion models,

    R. Bradbury and D. Zhong, “Your latent mask is wrong: Pixel-equivalent latent compositing for diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 18 630–18 639

    2026

  55. [55]

    Scale space diffusion,

    S. Mukhopadhyay, P. Udhayanan, and A. Shrivastava, “Scale space diffusion,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2026, pp. 35 851–35 860

    2026

  56. [56]

    Pixel motion diffusion is what we need for robot control,

    E.-R. Nguyen, Y . Zhang, K. Ranasinghe, X. Li, and M. S. Ryoo, “Pixel motion diffusion is what we need for robot control,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 23 663– 23 672

    2026

  57. [57]

    Simpler diffusion: 1.5 fid on imagenet512 with pixel-space diffusion,

    E. Hoogeboom, T. Mensink, J. Heek, K. Lamerigts, R. Gao, and T. Salimans, “Simpler diffusion: 1.5 fid on imagenet512 with pixel-space diffusion,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 18 062– 18 071

    2025

  58. [58]

    Pixel-perfect depth with semantics-prompted diffusion transformers,

    G. Xu, H. Lin, H. Luo, X. Wang, J. Yao, L. Zhu, Y . Pu, C. Chi , H. Sun, B. Wanget al., “Pixel-perfect depth with semantics-prompted diffusion transformers,”Advances in Neu- ral Information Processing Systems, vol. 38, pp. 174 731– 174 755, 2026

    2026

  59. [59]

    Novel view synthesis with pixel-space diffu- sion models,

    N. Elata, B. Kawar, Y . Ostrovsky-Berman, M. Farber, and R. Sokolovsky, “Novel view synthesis with pixel-space diffu- sion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 26 756– 26 766

    2025

  60. [60]

    Pixel-space post-training of latent diffusion models,

    C. Zhang, S. Motwani, M. Yu, J. Hou, F. Juefei-Xu, S. Tsai, P. Vajda, Z. He, and J. Wang, “Pixel-space post-training of latent diffusion models,”arXiv preprint arXiv:2409.17565, 2024

    Pith/arXiv arXiv 2024

  61. [61]

    Back to basics: Let denoising generative models denoise,

    T. Li and K. He, “Back to basics: Let denoising generative models denoise,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2026, pp. 36 115–36 125

    2026

  62. [62]

    Hidream-o1-image: A natively unified image generative foundation model with pixel-level unified transformer,

    Q. Cai, J. Chen, C. Gao, Z. Gong, Y . Li, Y . Pan, Y . Peng, Z. Qiu, K. Yu, Y . Zhanget al., “Hidream-o1-image: A natively unified image generative foundation model with pixel-level unified transformer,”arXiv preprint arXiv:2605.11061, 2026

    Pith/arXiv arXiv 2026

  63. [63]

    Show- o: One single transformer to unify multimodal understanding and generation,

    J. Xie, W. Mao, Z. Bai, D. J. Zhang, W. Wang, K. Q. Lin, Y . Gu, Z. Chen, Z. Yang, and M. Z. Shou, “Show- o: One single transformer to unify multimodal understanding and generation,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 28 240–28 264

    2025

  64. [64]

    Transfusion: Predict the next token and diffuse images with one multi-modal model,

    C. Zhou, L. Yu, A. Babu, K. Tirumala, M. Yasunaga, L. Shamis, J. Kahn, X. Ma, L. Zettlemoyer, and O. Levy, “Transfusion: Predict the next token and diffuse images with one multi-modal model,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 6446–6469

    2025

  65. [65]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  66. [66]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El- Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  67. [67]

    simple diffusion: End-to-end diffusion for high resolution images,

    E. Hoogeboom, J. Heek, and T. Salimans, “simple diffusion: End-to-end diffusion for high resolution images,” inInterna- tional Conference on Machine Learning. PMLR, 2023, pp. 13 213–13 232

    2023

  68. [68]

    Low light image enhancement challenge at ntire 2026,

    G. Ciubotariu, A. Rehman, F. A. Dharejo, R. A. Naqvi, M. V . Conde, R. Timofte, Z. Jin, H. Wu, W. Zhang, C. Yeet al., “Low light image enhancement challenge at ntire 2026,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 1741–1753

    2026

  69. [69]

    Hyperspec- tral imaging,

    D. Hong, C. Li, N. Yokoya, B. Zhang, X. Jia, A. Plaza, P. Gamba, J. A. Benediktsson, and J. Chanussot, “Hyperspec- tral imaging,”Nature Reviews Methods Primers, vol. 6, no. 1, p. 19, 2026

    2026

  70. [70]

    Perception encoder: The best visual embeddings are not at the output of the network,

    D. Bolya, P.-Y . Huang, P. Sun, J. H. Cho, A. Madotto, C. Wei, T. Ma, J. Zhi, J. Rajasegaran, H. Bangalathet al., “Perception encoder: The best visual embeddings are not at the output of the network,”Advances in Neural Information Processing Systems, vol. 38, pp. 60 884–60 937, 2026

    2026

  71. [71]

    Ntire 2026 challenge on video saliency prediction: Methods and results,

    A. Moskalenko, A. Bryncev, I. Kosmynin, K. Shilovskaya, M. Erofeev, D. Vatolin, R. Timofte, K. Wang, Y . Hu, Z. Li et al., “Ntire 2026 challenge on video saliency prediction: Methods and results,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2026, pp. 2193–2206

    2026

  72. [72]

    Ntire 2026 challenge on robust ai-generated image detection in the wild,

    A. Gushchin, K. Abud, E. Shumitskaya, A. Filippov, G. By- chkov, S. Lavrushkin, M. Erofeev, A. Antsiferova, C. Chen, S. Tanet al., “Ntire 2026 challenge on robust ai-generated image detection in the wild,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 1895–1913

    2026

  73. [73]

    One step diffusion via shortcut models,

    K. Frans, D. Hafner, S. Levine, and P. Abbeel, “One step diffusion via shortcut models,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 34 668–34 684

    2025

  74. [74]

    Latent diffusion for language generation,

    J. Lovelace, V . Kishore, C. Wan, E. Shekhtman, and K. Q. Weinberger, “Latent diffusion for language generation,”Ad- vances in Neural Information Processing Systems, vol. 36, pp. 56 998–57 025, 2023

    2023

  75. [75]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  76. [76]

    Turbo-vaed: Fast and stable transfer of video-vaes to mobile devices,

    Y . Zou, J. Yao, S. Yu, S. Zhang, W. Liu, and X. Wang, “Turbo-vaed: Fast and stable transfer of video-vaes to mobile devices,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 16, 2026, pp. 14 086–14 094

    2026

  77. [77]

    Sparc3d: Sparse representation and construction for high-resolution 3d shapes modeling,

    Z. Li, Y . Wang, H. Zheng, Y . Luo, and B. Wen, “Sparc3d: Sparse representation and construction for high-resolution 3d shapes modeling,”Advances in Neural Information Processing Systems, vol. 38, pp. 118 582–118 600, 2026

    2026

  78. [78]

    Vision foundation models can be good tokenizers for latent diffusion models,

    T. Bi, X. Zhang, Y . Lu, and N. Zheng, “Vision foundation models can be good tokenizers for latent diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 43 310–43 319

    2026

  79. [79]

    Softvq-vae: Efficient 1- dimensional continuous tokenizer,

    H. Chen, Z. Wang, X. Li, X. Sun, F. Chen, J. Liu, J. Wang, B. Raj, Z. Liu, and E. Barsoum, “Softvq-vae: Efficient 1- dimensional continuous tokenizer,” inProceedings of the Com- puter Vision and Pattern Recognition Conference, 2025, pp. 28 358–28 370

    2025

  80. [80]

    Mgvq: Could vq-vae beat vae? a generalizable tokenizer with multi-group quantization,

    M. Jia, W. Yin, X. Hu, J. Guo, X. Guo, Q. Zhang, X.- X. Long, and P. Tan, “Mgvq: Could vq-vae beat vae? a generalizable tokenizer with multi-group quantization,”arXiv preprint arXiv:2507.07997, 2025

    Pith/arXiv arXiv 2025

Showing first 80 references.