pith. sign in

arxiv: 2606.19970 · v1 · pith:WY53BIP5new · submitted 2026-06-18 · 💻 cs.CV

CrossFlow: One-Step Generation Across Latent and Pixel Spaces

Xiyuan Wang , Xiao Zhang , Yang Li , Ruoxi Jiang , Zhao Zhong , Liefeng Bo , Muhan Zhang This is my paper

Pith reviewed 2026-06-26 18:30 UTC · model grok-4.3

classification 💻 cs.CV
keywords CrossFlowone-step generationlatent to pixelflow matchingImageNet FIDdecoder replacementcross-space objective
0
0 comments X

The pith

CrossFlow maps noisy latent inputs directly to pixel images in one step.

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

The paper presents CrossFlow as a flow formulation that moves the probability path into latent space while supervising predictions with full pixel images. This setup eliminates the need for a separately trained decoder whose inputs may not match the clean latents seen during its own training. The resulting model reaches 1.62 FID on class-conditional ImageNet-1k at 256 by 256 resolution using only one function evaluation. It can operate either as a standalone one-step generator or as a drop-in decoder replacement inside existing latent diffusion pipelines. Ablations indicate that the latent encoder together with pixel-space perceptual and adversarial losses are required to reach the reported fidelity.

Core claim

CrossFlow defines a velocity-free one-step objective in which the latent trajectory supplies the training path while the supervised target is a pixel-space image rather than a latent displacement, allowing a single network to generate directly from noisy latents to pixels and to replace the decoder in latent diffusion pipelines.

What carries the argument

velocity-free one-step objective that uses the latent trajectory for the path but supervises pixel-image prediction

If this is right

  • One trained network replaces both the latent-space generator and the separate decoder at inference time.
  • Class-conditional ImageNet-1k 256 by 256 generation reaches 1.62 FID with a single function evaluation.
  • Pixel perceptual and adversarial losses, when paired with the latent encoder, become essential for maintaining output fidelity.
  • The same cross-space objective can be inserted into existing latent diffusion pipelines without retraining the upstream components.

Where Pith is reading between the lines

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

  • Training pipelines could drop the separate decoder stage entirely, reducing the number of models that must be optimized and stored.
  • Direct pixel-space supervision during flow training might allow perceptual metrics to influence the generative path more tightly than post-hoc decoding permits.
  • The formulation could be tested on video or audio by swapping the latent encoder for a modality-specific compressor while keeping the pixel-level (or waveform-level) supervision.

Load-bearing premise

The latent trajectory can define the training path while the supervised target remains a full pixel image without introducing new mismatches that need extra correction terms.

What would settle it

Running a standard latent diffusion model plus its trained decoder on the same one-step budget and showing that its FID on ImageNet-1k 256 by 256 exceeds 1.62 while CrossFlow outputs exhibit visible artifacts or lower perceptual scores.

Figures

Figures reproduced from arXiv: 2606.19970 by Liefeng Bo, Muhan Zhang, Ruoxi Jiang, Xiao Zhang, Xiyuan Wang, Yang Li, Zhao Zhong.

Figure 1
Figure 1. Figure 1: CrossFlow generation paradigm. Latent diffusion uses a two-stage pipeline: iterative denoising in latent space followed by VAE decoding. Pixel-space diffusion performs iterative denoising directly in the image domain. CrossFlow uses a single model to map a noisy latent prior directly to a pixel-space image, unifying one-step generation and latent-to-pixel decoding. We propose CrossFlow, a cross-space one-s… view at source ↗
Figure 2
Figure 2. Figure 2: Uncurated 1-NFE class-conditional samples on ImageNet 256 × 256. Samples are generated by CrossFlow-XL from latent noise directly to pixels. Rows correspond to class 12 (house finch, linnet, Carpodacus mexicanus), class 309 (bee), class 698 (palace), and class 973 (coral reef), respectively. tuple represents the corresponding tangent directions. To optimize computational efficiency, we execute this step us… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of Fθ(zt, t, 0) as t varies from 0 (left, clean latent) to 1 (right, pure noise). Each row represents a different semantic category. The interpolation preserves recognizable structure while progressively introducing stochastic variation and noise. We implement Fθ with a Vision Transformer (ViT) backbone [40]. Following ViT conventions, we evaluate three variants: CrossFlow-B with 12 layers an… view at source ↗
Figure 5
Figure 5. Figure 5: CrossFlow as a VAE decoder. We evaluate CrossFlow as a VAE decoder for latent diffusion and report FID over training epochs for a LightningDiT-B/1 generator in the VA-VAE latent space. CrossFlow consistently improves generation quality over the VA-VAE decoder. 4.4 Performance as a VAE Decoder Finally, we evaluate CrossFlow as a decoder in a VAE-style pipeline. This setting tests whether the same latent-to-… view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of GAN collapse under a vanilla adversarial loss. Rows show different tuples [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Gradient norms of the GAN loss and CrossFlow loss across time, averaged over 204,800 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
read the original abstract

Most diffusion and flow-matching generators define the prior, probability path, and prediction target in the same representation space. Latent diffusion improves efficiency by moving this path into an autoencoder latent space, but the final sample is still produced by a separately trained decoder. This separation creates a mismatch: the generator is optimized for latent-space prediction, while final quality depends on how the decoder handles generated latents that may differ from clean encoder outputs. We introduce CrossFlow, a cross-space flow formulation that maps noisy latent inputs directly to pixel-space images. The key technical step is a velocity-free one-step objective: the latent trajectory defines the training path, but the supervised prediction is an image rather than a latent displacement. This lets one model act both as a one-step latent-to-pixel generator and as a decoder replacement for latent diffusion pipelines. On class-conditional ImageNet-1k at $256\times256$, CrossFlow-XL achieves 1.62 FID with one function evaluation. Ablations show that the latent encoder and pixel-space perceptual and adversarial losses are important for fidelity. These results indicate that cross-space flow objectives can combine the efficiency of latent representations with direct pixel-space supervision, without requiring a separate decoder at inference.

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

1 major / 2 minor

Summary. The manuscript introduces CrossFlow, a cross-space flow model that maps noisy latent inputs directly to pixel-space images via a velocity-free one-step objective. The latent encoder trajectory defines the training path while the supervised target is a full pixel image rather than a latent displacement. This design is claimed to let a single model serve as both a one-step latent-to-pixel generator and a decoder replacement for latent diffusion pipelines. On class-conditional ImageNet-1k at 256×256, CrossFlow-XL reports 1.62 FID with one function evaluation; ablations indicate that the latent encoder and pixel-space perceptual/adversarial losses are important.

Significance. If the alignment between training and inference latent distributions holds without additional correction terms, the approach would combine latent-space efficiency with direct pixel supervision and eliminate the separate decoder stage, which is a meaningful simplification for generative pipelines. The reported one-step FID is competitive and the dual-use capability would be a clear strength if demonstrated rigorously.

major comments (1)
  1. [Abstract / §3] Abstract and §3 (method): the velocity-free one-step objective trains on paths defined by the latent encoder but supervises directly on pixel images. For the model to function as a decoder replacement at inference, generated latents must lie on the same distribution as the encoder outputs used during training. No derivation, alignment proof, or quantitative analysis is provided showing why this cross-space mapping preserves the required distribution without introducing mismatches that would necessitate correction terms or multi-step refinement—the central claim of the paper.
minor comments (2)
  1. [Results] The abstract reports an FID value and ablation importance but provides no error bars, full experimental protocol, training details, or comparison tables; these should be added to the results section for reproducibility.
  2. [§3] Notation for the velocity-free objective and the cross-space mapping should be formalized with explicit equations rather than descriptive text only.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the single major comment below and outline revisions to improve clarity on the distribution alignment aspect of the method.

read point-by-point responses

  1. Referee: [Abstract / §3] Abstract and §3 (method): the velocity-free one-step objective trains on paths defined by the latent encoder but supervises directly on pixel images. For the model to function as a decoder replacement at inference, generated latents must lie on the same distribution as the encoder outputs used during training. No derivation, alignment proof, or quantitative analysis is provided showing why this cross-space mapping preserves the required distribution without introducing mismatches that would necessitate correction terms or multi-step refinement—the central claim of the paper.

    Authors: We agree that the manuscript provides no formal derivation or theoretical proof of distribution alignment between training and inference latents. The work is primarily empirical: the velocity-free objective uses the encoder trajectory to define the input noise path while supervising on pixel targets, and the reported results (1.62 FID in one step) demonstrate that the trained model produces high-quality outputs when applied to latents drawn from the same encoder distribution. We do not claim a general guarantee that mismatches are always absent; rather, the design and pixel-space losses are intended to make the mapping robust in practice. To address the comment, we will revise §3 to add an explicit discussion of this assumption and include new quantitative analysis (e.g., measuring latent-space statistics or reconstruction error on generated vs. encoder latents) in the experiments section. This revision will clarify the empirical basis without overstating theoretical guarantees. revision: yes

Circularity Check

0 steps flagged

No circularity: cross-space objective is independently defined

full rationale

The paper introduces a velocity-free one-step objective where the training path is defined by the latent encoder trajectory but the supervised target is a pixel image. This formulation is presented as a new technical step without any equations or claims that reduce the reported FID result or the model's dual role to a fitted parameter, self-citation chain, or input by construction. No self-citations are invoked as load-bearing uniqueness theorems, no ansatz is smuggled via prior work, and no renaming of known results occurs. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the approach relies on standard autoencoder latents and perceptual/adversarial losses whose details are not specified here.

pith-pipeline@v0.9.1-grok · 5758 in / 1017 out tokens · 29458 ms · 2026-06-26T18:30:04.958553+00:00 · methodology

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

58 extracted references · 4 linked inside Pith

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inNeurIPS, 2020

    2020

  2. [2]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” inICLR, 2021

    2021

  3. [3]

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

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,”ICLR, 2023

    2023

  4. [4]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, and M. Nickel, “Flow matching for generative modeling,”ICLR, 2023

    2023

  5. [5]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,”ICCV, 2023

    2023

  6. [6]

    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,”ECCV, 2024

    2024

  7. [7]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esser, S. K. andreas Blattmann, R. Entezari, J. Müller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, and et al, “Scaling rectified flow transformers for high-resolution image synthesis,”ICML, 2024

    2024

  8. [8]

    B. F. Labs, “Flux,” https://github.com/black-forest-labs/flux, 2024

    2024

  9. [9]

    Sana: Efficient high-resolution image synthesis with linear diffusion transformers,

    E. Xie, J. Chen, J. Chen, H. Cai, Y . Lin, Z. Zhang, M. Li, Y . Lu, and S. Han, “Sana: Efficient high-resolution image synthesis with linear diffusion transformers,”ICLR, 2025

    2025

  10. [10]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inICLR, 2021

    2021

  11. [11]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,”ICML, 2023

    2023

  12. [12]

    Improved techniques for training consistency models,

    Y . Song and P. Dhariwal, “Improved techniques for training consistency models,”ICML, 2024

    2024

  13. [13]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,”ICLR, 2022

    2022

  14. [14]

    One-step diffusion with distribution matching distillation,

    T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Freeman, and T. Park, “One-step diffusion with distribution matching distillation,”CVPR, 2024

    2024

  15. [15]

    Mean flows for one-step generative modeling,

    Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He, “Mean flows for one-step generative modeling,” NeurIPS, 2025

    2025

  16. [16]

    Improved mean flows: On the challenges of fastforward generative models, 2025b,

    Z. Geng, Y . Lu, Z. Wu, E. Shechtman, J. Z. Kolter, and K. He, “Improved mean flows: On the challenges of fastforward generative models, 2025b,” 2026

    2026

  17. [17]

    High-resolution image synthesis with latent diffusion models,

    R. R. andreas Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,”CVPR, 2022

    2022

  18. [18]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Q. Nichol, “Diffusion models beat gans on image synthesis,” inNeurIPS, 2021

    2021

  19. [19]

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

    D. Podell, Z. English, K. L. andreas Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach, “Sdxl: Improving latent diffusion models for high-resolution image synthesis,”ICLR, 2023

    2023

  20. [20]

    Neural discrete representation learning,

    A. V . D. Oord, O. Vinyals, and et al, “Neural discrete representation learning,”NeurIPS, 2017

    2017

  21. [21]

    Taming transformers for high-resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high-resolution image synthesis,”CVPR, 2021

    2021

  22. [22]

    Deep compression autoencoder for efficient high-resolution diffusion models,

    J. Chen, H. Cai, J. Chen, E. Xie, S. Yang, H. Tang, M. Li, Y . Lu, and S. Han, “Deep compression autoencoder for efficient high-resolution diffusion models,”ICLR, 2025

    2025

  23. [23]

    Score-based generative modeling in latent space,

    A. Vahdat, K. Kreis, and J. Kautz, “Score-based generative modeling in latent space,” 2021

    2021

  24. [24]

    Repa-e: Unlocking vae for end-to-end tuning with latent diffusion transformers,

    X. Leng, J. Singh, Y . Hou, Z. Xing, S. Xie, and L. Zheng, “Repa-e: Unlocking vae for end-to-end tuning with latent diffusion transformers,”ICCV, 2025

    2025

  25. [25]

    Diffusion as self-distillation: End-to-end latent diffusion in one model,

    X. Wang and M. Zhang, “Diffusion as self-distillation: End-to-end latent diffusion in one model,”CoRR, vol. abs/2511.14716, 2025

    arXiv 2025

  26. [26]

    Unified latents (ul): How to train your latents,

    J. Heek, E. Hoogeboom, T. Mensink, and T. Salimans, “Unified latents (ul): How to train your latents,” arXiv:2602.17270, 2026. 11

    arXiv 2026

  27. [27]

    End-to-end training for unified tokenization and latent denoising,

    S. Duggal, X. Bai, Z. Wu, R. Zhang, E. Shechtman, A. Torralba, P. Isola, and W. T. Freeman, “End-to-end training for unified tokenization and latent denoising,”CoRR, vol. abs/2603.22283, 2026

    arXiv 2026

  28. [28]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. C. Courville, and Y . Bengio, “Generative adversarial nets,”NeurIPS, 2014

    2014

  29. [29]

    Large scale GAN training for high fidelity natural image synthesis,

    A. Brock, J. Donahue, and K. Simonyan, “Large scale GAN training for high fidelity natural image synthesis,” inICLR, 2019

    2019

  30. [30]

    StyleGAN-XL: Scaling StyleGAN to large diverse datasets,

    A. Sauer, K. Schwarz, and A. Geiger, “StyleGAN-XL: Scaling StyleGAN to large diverse datasets,” in SIGGRAPH, 2022

    2022

  31. [31]

    Variational inference with normalizing flows,

    D. J. Rezende and S. Mohamed, “Variational inference with normalizing flows,”ICML, 2015

    2015

  32. [32]

    Density estimation using real nvp,

    L. Dinh, J. Sohl-Dickstein, and S. Bengio, “Density estimation using real nvp,”ICLR, 2017

    2017

  33. [33]

    Augmented normalizing flows: Bridging the gap between generative flows and latent variable models,

    C. Huang, L. Dinh, and A. C. Courville, “Augmented normalizing flows: Bridging the gap between generative flows and latent variable models,”ICML, 2020

    2020

  34. [34]

    Lifting architectural constraints of injective flows,

    P. Sorrenson, F. Roth, K. Dreczkowski, V . Stimper, and F. Noé, “Lifting architectural constraints of injective flows,”ICLR, 2024

    2024

  35. [35]

    Consistency trajectory models: Learning probability flow ODE trajectory of diffusion,

    D. Kim, C. Lai, W. Liao, N. Murata, Y . Takida, T. Uesaka, Y . He, Y . Mitsufuji, and S. Ermon, “Consistency trajectory models: Learning probability flow ODE trajectory of diffusion,” inICLR, 2024

    2024

  36. [36]

    One step diffusion via shortcut models,

    K. Frans, D. Hafner, S. Levine, and P. Abbeel, “One step diffusion via shortcut models,” inICLR, 2025

    2025

  37. [37]

    Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models,

    J. Yao, B. Yang, and X. Wang, “Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models,” inCVPR, 2025

    2025

  38. [38]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in2009 IEEE conference on computer vision and pattern recognition. IEEE, 2009, pp. 248–255

    2009

  39. [39]

    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,”Advances in neural information processing systems, vol. 30, 2017

    2017

  40. [40]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Min- derer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inICLR, 2021

    2021

  41. [41]

    Muon: An optimizer for hidden layers in neural networks,

    K. Jordan, Y . Jin, V . Boza, Y . Jiacheng, F. Cesista, L. Newhouse, and J. Bernstein, “Muon: An optimizer for hidden layers in neural networks,” 2024. [Online]. Available: https://kellerjordan.github.io/posts/muon/

    2024

  42. [42]

    Muon is scalable for LLM training,

    J. Liu, J. Su, X. Yao, Z. Jiang, G. Lai, Y . Du, Y . Qin, W. Xu, E. Lu, J. Yan, Y . Chen, H. Zheng, Y . Liu, S. Liu, B. Yin, W. He, H. Zhu, Y . Wang, J. Wang, M. Dong, Z. Zhang, Y . Kang, H. Zhang, X. Xu, Y . Zhang, Y . Wu, X. Zhou, and Z. Yang, “Muon is scalable for LLM training,”Arxiv abs/2502.16982, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Representation alignment for generation: Training diffusion transformers is easier than you think,

    S. Yu, S. Kwak, H. Jang, J. Jeong, J. Huang, J. Shin, and S. Xie, “Representation alignment for generation: Training diffusion transformers is easier than you think,”ICLR, 2025

    2025

  44. [44]

    Scaling up GANs for text-to-image synthesis,

    M. Kang, J.-Y . Zhu, R. Zhang, J. Park, E. Shechtman, S. Paris, and T. Park, “Scaling up GANs for text-to-image synthesis,” inCVPR, 2023

    2023

  45. [45]

    There is no V AE: End-to-end pixel-space generative modeling via self-supervised pre-training,

    J. Lei, K. Liu, J. Berner, H. Yu, H. Zheng, J. Wu, and X. Chu, “There is no V AE: End-to-end pixel-space generative modeling via self-supervised pre-training,” inICLR, 2026

    2026

  46. [46]

    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,” inICML, 2023

    2023

  47. [47]

    One-step latent-free image generation with pixel mean flows,

    Y . Lu, S. Lu, Q. Sun, H. Zhao, Z. Jiang, X. Wang, T. Li, Z. Geng, and K. He, “One-step latent-free image generation with pixel mean flows,”arXiv:2601.22158, 2026

    Pith/arXiv arXiv 2026

  48. [48]

    Understanding diffusion objectives as the ELBO with simple data augmentation,

    D. P. Kingma and R. Gao, “Understanding diffusion objectives as the ELBO with simple data augmentation,” inNeurIPS, 2023

    2023

  49. [49]

    Simpler diffusion (SiD2): 1.5 FID on ImageNet512 with pixel-space diffusion,

    E. Hoogeboom, T. Mensink, J. Heek, K. Lamerigts, R. Gao, and T. Salimans, “Simpler diffusion (SiD2): 1.5 FID on ImageNet512 with pixel-space diffusion,” inCVPR, 2025. 12

    2025

  50. [50]

    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,”arXiv preprint arXiv:2511.20645, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    FLUX.2: Frontier Visual Intelligence,

    B. F. Labs, “FLUX.2: Frontier Visual Intelligence,” https://bfl.ai/blog/flux-2, 2025

    2025

  52. [52]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, high-performance deep learning library,” in NeurIPS, 2019

    2019

  53. [53]

    Transformers: State-of-the-art natural language processing,

    T. Wolf, L. Debut, V . Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y . Jernite, J. Plu, C. Xu, T. L. Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush, “Transformers: State-of-the-art natural language processing,” inEMNLP, 2020

    2020

  54. [54]

    Pytorch image models,

    R. Wightman, “Pytorch image models,” 2019

    2019

  55. [55]

    High-fidelity performance metrics for generative models in pytorch,

    A. Obukhov, M. Seitzer, P.-W. Wu, S. Zhydenko, J. Kyl, and E. Y .-J. Lin, “High-fidelity performance metrics for generative models in pytorch,” 2020. [Online]. Available: https://github.com/toshas/torch-fidelity

    2020

  56. [56]

    Is noise conditioning necessary for denoising generative models?

    Q. Sun, Z. Jiang, H. Zhao, and K. He, “Is noise conditioning necessary for denoising generative models?” inICML, 2025

    2025

  57. [57]

    Diffusion transformers with representation autoencoders,

    B. Zheng, N. Ma, S. Tong, and S. Xie, “Diffusion transformers with representation autoencoders,”ICLR, 2026

    2026

  58. [58]

    Siméoni, H

    O. Siméoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. E. Yi, M. Ramamonjisoa, F. Massa, D. Haziza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H. Jégou, P. Labatut, and P. Bojanowski, “Dinov3,”Arxiv abs/2508.10104, 2025. 13...

    Pith/arXiv arXiv 2025