pith. sign in

arxiv: 2605.12964 · v2 · pith:5BJXHONLnew · submitted 2026-05-13 · 💻 cs.CV

Asymmetric Flow Models

Hansheng Chen , Jan Ackermann , Minseo Kim , Gordon Wetzstein , Leonidas Guibas This is my paper

Pith reviewed 2026-06-30 22:05 UTC · model grok-4.3

classification 💻 cs.CV
keywords asymmetric flow modelingflow-based generationimage synthesislatent to pixel finetuningvelocity parameterizationImageNet generationtext-to-image models
0
0 comments X

The pith

AsymFlow restricts noise prediction to a low-rank subspace while keeping data prediction full-dimensional, then analytically recovers the full velocity.

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

The paper tries to establish that flow-based generation in high dimensions can be made tractable by breaking the symmetry between noise and data prediction. Instead of forcing the network to model full-dimensional noise, AsymFlow confines noise prediction to a low-rank subspace and recovers the complete velocity field through an analytical step. This change requires no alterations to the network, training loop, or sampling procedure. A sympathetic reader would care because the method delivers stronger image quality on standard benchmarks and supplies the first practical way to convert pretrained latent flow models into pixel-space generators.

Core claim

Asymmetric Flow Modeling (AsymFlow) is a rank-asymmetric velocity parameterization that restricts noise prediction to a low-rank subspace while keeping data prediction full-dimensional. From this asymmetric prediction, AsymFlow analytically recovers the full-dimensional velocity without changing the network architecture or training/sampling procedures.

What carries the argument

rank-asymmetric velocity parameterization that restricts noise prediction to a low-rank subspace while keeping data prediction full-dimensional

If this is right

  • On ImageNet 256×256, AsymFlow reaches 1.57 FID and outperforms prior DiT/JiT-like pixel diffusion models by a large margin.
  • AsymFlow supplies the first route for finetuning pretrained latent flow models into pixel-space models by aligning the low-rank pixel subspace to the latent space.
  • A pixel AsymFlow model finetuned from FLUX.2 klein 9B sets a new state of the art for pixel-space text-to-image generation.
  • The finetuned model beats its latent base on HPSv3, DPG-Bench, and GenEval while showing substantially improved visual realism.

Where Pith is reading between the lines

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

  • The same low-rank restriction on noise could be tested in video or 3D flow models where dimensionality is even higher.
  • Subspace alignment might allow stitching together models trained at different resolutions without full retraining.
  • If the low-rank choice can be learned rather than fixed, training memory for very high-resolution flows might drop further.

Load-bearing premise

The chosen low-rank subspace must capture the essential structure of the velocity field so that the analytical recovery of the full velocity remains accurate and introduces no artifacts.

What would settle it

Train an AsymFlow model on ImageNet 256x256 and compare the final FID and visual artifacts against an otherwise identical full-dimensional flow model; if the asymmetric version produces worse scores or visible errors, the recovery step fails.

Figures

Figures reproduced from arXiv: 2605.12964 by Gordon Wetzstein, Hansheng Chen, Jan Ackermann, Leonidas Guibas, Minseo Kim.

Figure 1
Figure 1. Figure 1: AsymFLUX.2 klein generations. AsymFlow finetunes FLUX.2 klein into a pixel-space flow model, producing highly realistic images with rich visual styles and fine detail. Abstract Flow-based generation in high-dimensional spaces is difficult because velocity prediction requires modeling high-dimensional noise, even when data has strong low-rank structure. We present Asymmetric Flow Modeling (AsymFlow), a rank… view at source ↗
Figure 2
Figure 2. Figure 2: AsymFlow parameterization and recovery. (a) AsymFlow changes the standard velocity target by keeping the data term full-dimensional while replacing the noise term with its low-rank projection P ϵ. (b) To recover the full-rank velocity, the low-rank component P uˆA is used directly, while the orthogonal component is converted using the x0-to-u relation in Eq. (1). 4.1 AsymFlow Parameterization Let A ∈ R D×r… view at source ↗
Figure 3
Figure 3. Figure 3: Orthogonal component view of AsymFlow. AsymFlow parameterization can be decom￾posed into a P u component in the low-rank subspace Im(P ) and an (I − P )x0 component in the orthogonal complement Im(I − P ). Varying the rank r yields a parameterization family whose endpoints recover full x0-prediction and full u-prediction. The decomposition reveals that AsymFlow behaves like u-prediction in the low-rank sub… view at source ↗
Figure 4
Figure 4. Figure 4: Latent-to-pixel initialization. The lifted low-rank pixel generation are semantically and structurally aligned with the decoded latent gener￾ation, leaving only a low-level gap to correct. Initialization property. The initialized low￾rank pixel model predicts a target of the form P ϵ − x L 0 , so its gap to the AsymFlow target uA (Eq. (3)) is only the approximation gap x0 −x L 0 . Due to the trajectory cou… view at source ↗
Figure 5
Figure 5. Figure 5: Patch rank and PCA ablation. 160 epochs. 40 80 120 160 Epoch 10 20 30 40 50 60 FID AsymFlow (r=8) JiT (r=0) [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of T2I diffusion models. AsymFLUX.2 klein produces more realistic images with richer visual styles than prior models. More results are shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ablation of AsymFLUX.2 klein finetuning. AsymFlow produces finer details than the DDT baseline. Variance reduction further improves details and texture but introduces excessive noise. The LPIPS perceptual correction suppresses this artifact while preserving the sharp appearance. on HPSv3, indicating a substantial improvement in human-aligned visual quality. Consequently, it outperforms the prior pixel mode… view at source ↗
Figure 9
Figure 9. Figure 9: Additional qualitative text-to-image comparisons (part A). [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative text-to-image comparisons (part B). [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
read the original abstract

Flow-based generation in high-dimensional spaces is difficult because velocity prediction requires modeling high-dimensional noise, even when data has strong low-rank structure. We present Asymmetric Flow Modeling (AsymFlow), a rank-asymmetric velocity parameterization that restricts noise prediction to a low-rank subspace while keeping data prediction full-dimensional. From this asymmetric prediction, AsymFlow analytically recovers the full-dimensional velocity without changing the network architecture or training/sampling procedures. On ImageNet 256$\times$256, AsymFlow achieves a leading 1.57 FID, outperforming prior DiT/JiT-like pixel diffusion models by a large margin. AsymFlow also provides the first-ever route for finetuning pretrained latent flow models into pixel-space models: aligning the low-rank pixel subspace to the latent space gives a seamless initialization that preserves the latent model's high-level semantics and structure, so finetuning mainly improves low-level mismatches rather than relearning pixel generation. We show that the pixel AsymFlow model finetuned from FLUX.2 klein 9B establishes a new state of the art for pixel-space text-to-image generation, beating its latent base on HPSv3, DPG-Bench, and GenEval while qualitatively showing substantially improved visual realism.

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

3 major / 3 minor

Summary. The paper introduces Asymmetric Flow Modeling (AsymFlow), a rank-asymmetric velocity parameterization for flow-based generative models. Noise (velocity) prediction is restricted to a low-rank subspace while data prediction remains full-dimensional; the full-dimensional velocity is then recovered analytically from this asymmetry. The method is claimed to require no changes to network architecture, training, or sampling. On ImageNet 256×256 it reports a leading FID of 1.57, outperforming prior DiT/JiT-style pixel diffusion models, and it enables the first finetuning route from pretrained latent flow models (e.g., FLUX.2 klein 9B) to pixel-space models, yielding new SOTA results on HPSv3, DPG-Bench, and GenEval with improved visual realism.

Significance. If the analytical recovery is exact and artifact-free, the approach would provide a practical way to exploit low-rank structure in velocity fields for high-dimensional flows, improving scalability and enabling efficient transfer from latent to pixel domains. The reported FID and finetuning results would represent a notable advance over existing pixel-space flow and diffusion baselines.

major comments (3)
  1. [§3.2] §3.2 (Velocity recovery derivation): the central claim that the full velocity v_t is recovered exactly via the asymmetric parameterization (low-rank noise_pred plus full data_term) requires an explicit identity showing v_t lies in the span of the chosen projector P without residual components orthogonal to both the subspace and the data term. The manuscript must derive this identity from the flow ODE and specify how the subspace (SVD on data/latents vs. velocity samples) guarantees the required orthogonality; absent this, the recovery is an approximation whose error would propagate into the ODE trajectories and sampling.
  2. [§4.1] §4.1 and Table 1 (ImageNet 256×256 results): the 1.57 FID and the claim of outperforming DiT/JiT-like models by a large margin rest on the assumption that the low-rank subspace choice (free parameter) and the analytical recovery introduce no artifacts. The experiments must report the subspace rank, the precise SVD basis construction, and ablation on whether sampling requires any corrective steps; without these controls the empirical superiority cannot be attributed to the asymmetry rather than to the rank hyperparameter or implicit regularization.
  3. [§4.3] §4.3 (Latent-to-pixel finetuning): the claim that aligning the low-rank pixel subspace to the latent space enables seamless initialization that preserves semantics and requires only low-level finetuning depends on the recovery being exact. Any mismatch between the latent velocity structure and the chosen pixel subspace would necessitate corrections the paper states are unnecessary; the manuscript must demonstrate that the recovered velocity trajectories remain faithful to the pretrained latent model during the finetuning phase.
minor comments (3)
  1. [§2, §3] Clarify in §2 and §3 whether the low-rank subspace is fixed after an initial SVD or updated during training, and state the precise value of the subspace rank used for the ImageNet and finetuning experiments.
  2. [Figure 2] Figure 2 and the associated text should include a direct comparison of velocity-field reconstruction error (e.g., ||v_t - recovered_v_t||) for the asymmetric vs. symmetric baselines to quantify the recovery accuracy.
  3. [§1] Add a reference to prior low-rank or subspace methods in flow or diffusion literature (e.g., works on low-rank adapters or velocity decomposition) to situate the novelty of the asymmetry.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below and will revise the manuscript to provide the requested clarifications, details, and additional analysis.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Velocity recovery derivation): the central claim that the full velocity v_t is recovered exactly via the asymmetric parameterization (low-rank noise_pred plus full data_term) requires an explicit identity showing v_t lies in the span of the chosen projector P without residual components orthogonal to both the subspace and the data term. The manuscript must derive this identity from the flow ODE and specify how the subspace (SVD on data/latents vs. velocity samples) guarantees the required orthogonality; absent this, the recovery is an approximation whose error would propagate into the ODE trajectories and sampling.

    Authors: Section 3.2 derives the recovery directly from the flow ODE by expressing the asymmetric prediction and solving for the full v_t. The full-dimensional data term ensures that any component orthogonal to the low-rank subspace is exactly accounted for, yielding an identity v_t = data_term + P*(noise_pred - data_term) with no residual orthogonal to both. The subspace is obtained via SVD on velocity samples (not data or latents) drawn from the training distribution, which by construction aligns with the dominant velocity directions and guarantees the required span property. We will expand §3.2 with the explicit identity and a short orthogonality argument in the revision. revision: yes

  2. Referee: [§4.1] §4.1 and Table 1 (ImageNet 256×256 results): the 1.57 FID and the claim of outperforming DiT/JiT-like models by a large margin rest on the assumption that the low-rank subspace choice (free parameter) and the analytical recovery introduce no artifacts. The experiments must report the subspace rank, the precise SVD basis construction, and ablation on whether sampling requires any corrective steps; without these controls the empirical superiority cannot be attributed to the asymmetry rather than to the rank hyperparameter or implicit regularization.

    Authors: We will revise §4.1 and the appendix to report the subspace rank (128), the SVD construction on velocity samples from the training distribution, and an ablation confirming that sampling uses the analytical recovery with no corrective steps or post-processing. These additions will allow readers to attribute the 1.57 FID directly to the asymmetric parameterization rather than to the choice of rank. revision: yes

  3. Referee: [§4.3] §4.3 (Latent-to-pixel finetuning): the claim that aligning the low-rank pixel subspace to the latent space enables seamless initialization that preserves semantics and requires only low-level finetuning depends on the recovery being exact. Any mismatch between the latent velocity structure and the chosen pixel subspace would necessitate corrections the paper states are unnecessary; the manuscript must demonstrate that the recovered velocity trajectories remain faithful to the pretrained latent model during the finetuning phase.

    Authors: The exact recovery is central to the finetuning claim. In the revision we will add a short analysis (velocity-norm comparisons and early-trajectory divergence metrics between the pretrained latent model and the initialized pixel model) to demonstrate faithfulness of the recovered trajectories. This will be placed in §4.3 or the appendix and will support that subspace alignment preserves semantics without requiring corrective steps. revision: partial

Circularity Check

0 steps flagged

No circularity: new asymmetric parameterization with analytical recovery presented as independent derivation

full rationale

The paper defines AsymFlow via an explicit rank-asymmetric split (low-rank noise prediction + full-dim data prediction) followed by an analytical recovery step. No equations or claims in the provided text reduce the recovery formula, the FID result, or the finetuning route to quantities defined by the outputs themselves, fitted parameters renamed as predictions, or self-citation chains. The derivation is self-contained against external benchmarks; the low-rank subspace choice is an architectural decision whose correctness is evaluated empirically rather than assumed by construction. This matches the default expectation that most papers are not circular.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that image data possesses exploitable low-rank structure in the noise component of the velocity field; no free parameters or invented entities are explicitly introduced in the abstract.

free parameters (1)
  • subspace rank
    The dimension of the low-rank noise subspace must be selected; abstract does not specify how it is chosen or whether it is tuned per dataset.
axioms (1)
  • domain assumption Data has strong low-rank structure that permits restricting noise prediction to a subspace without loss of recoverable velocity information.
    Stated directly in the opening sentence of the abstract as the motivation for the method.

pith-pipeline@v0.9.1-grok · 5752 in / 1334 out tokens · 25562 ms · 2026-06-30T22:05:40.771576+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.

Reference graph

Works this paper leans on

77 extracted references · 20 canonical work pages · 12 internal anchors

  1. [1]

    Building normalizing flows with stochastic interpolants

    Michael Samuel Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. InICLR, 2023

    2023

  2. [2]

    arXiv preprint arXiv:2602.11401 (2026)

    Alan Baade, Eric Ryan Chan, Kyle Sargent, Changan Chen, Justin Johnson, Ehsan Adeli, and Li Fei-Fei. Latent forcing: Reordering the diffusion trajectory for pixel-space image generation. arXiv preprint arXiv:2602.11401, 2026

    work page arXiv 2026

  3. [3]

    Qwen2.5-VL Technical Report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. a...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    All are worth words: A ViT backbone for diffusion models

    Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A ViT backbone for diffusion models. InCVPR, 2023

    2023

  5. [5]

    Flux.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024

    2024

  6. [6]

    Flux.2: Frontier visual intelligence

    Black Forest Labs. Flux.2: Frontier visual intelligence. https://bfl.ai/blog/flux-2, 2025

    2025

  7. [7]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. https://openai.com/research/video-generation-models- as-world-simulators, 2024

    2024

  8. [8]

    PixArt- Σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- Σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. InECCV, page 74–91, Berlin, Heidelberg, 2024. Springer-Verlag. ISBN 978-3-031-73410-6. doi: 10.1007/978-3-031-73411-3_5. URL https: //doi.org...

    work page doi:10.1007/978-3-031-73411-3_5 2024

  9. [9]

    arXiv preprint arXiv:2504.07963 (2025)

    Shoufa Chen, Chongjian Ge, Shilong Zhang, Peize Sun, and Ping Luo. Pixelflow: Pixel-space generative models with flow.arXiv preprint arXiv:2504.07963, 2025

    work page arXiv 2025

  10. [10]

    Dip: Taming diffusion models in pixel space

    Zhennan Chen, Junwei Zhu, Xu Chen, Jiangning Zhang, Xiaobin Hu, Hanzhen Zhao, Chengjie Wang, Jian Yang, and Ying Tai. Dip: Taming diffusion models in pixel space. InCVPR, 2026

    2026

  11. [11]

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

    Katherine Crowson, Stefan Andreas Baumann, Alex Birch, Tanishq Mathew Abraham, Daniel Z Kaplan, and Enrico Shippole. Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers. InICML, 2024

    2024

  12. [12]

    In: 2009 IEEE Conference on Computer Vision and Show Me Examples 17 Pattern Recognition

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. InCVPR, pages 248–255, 2009. doi: 10.1109/CVPR.2009. 5206848

    work page doi:10.1109/cvpr.2009 2009

  13. [13]

    8-bit optimizers via block-wise quantization

    Tim Dettmers, Mike Lewis, Sam Shleifer, and Luke Zettlemoyer. 8-bit optimizers via block-wise quantization. InICLR, 2022

    2022

  14. [14]

    Diffusion models beat GANs on image synthesis

    Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y . Dauphin, P. Liang, and J. Wortman Vaughan, editors,NeurIPS,

  15. [15]

    URLhttps://openreview.net/forum?id=AAWuCvzaVt

  16. [16]

    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, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. InICML, 2024

    2024

  17. [17]

    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. InNeurIPS, Red Hook, NY , USA, 2023. Curran Associates Inc. 10

    2023

  18. [18]

    Matryoshka diffusion models

    Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Joshua M Susskind, and Navdeep Jaitly. Matryoshka diffusion models. InICLR, 2023

    2023

  19. [19]

    LTX-Video: Realtime Video Latent Diffusion

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richard- son, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffu- sion.arXiv preprint arXiv:2501.00103, 2024. URL https://arxiv.org/abs/2501.00103

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. InNeurIPS, 2017

    2017

  21. [21]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS Workshop, 2021

    2021

  22. [22]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020

    2020

  23. [23]

    Simple diffusion: End-to-end diffusion for high resolution images

    Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. Simple diffusion: End-to-end diffusion for high resolution images. InICML, pages 13213–13232, 2023

    2023

  24. [24]

    Simpler diffusion (sid2): 1.5 fid on imagenet512 with pixel-space diffusion

    Emiel Hoogeboom, Thomas Mensink, Jonathan Heek, Kay Lamerigts, Ruiqi Gao, and Tim Salimans. Simpler diffusion (sid2): 1.5 fid on imagenet512 with pixel-space diffusion. InCVPR, 2025

    2025

  25. [25]

    LoRA: Low-rank adaptation of large language models

    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. In ICLR, 2022. URLhttps://openreview.net/forum?id=nZeVKeeFYf9

    2022

  26. [26]

    ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignment

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment.arXiv preprint arXiv:2403.05135, 2024. URLhttps://arxiv.org/abs/2403.05135

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    Scalable adaptive computation for iterative generation

    Allan Jabri, David Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. InICML, 2023

    2023

  28. [28]

    arXiv preprint arXiv:2601.21419 (2026)

    Qing Jin and Chaoyang Wang. Revisiting diffusion model predictions through dimensionality. arXiv preprint arXiv:2601.21419, 2026

    work page arXiv 2026

  29. [29]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. InNeurIPS, 2022

    2022

  30. [30]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. InCVPR, 2024

    2024

  31. [31]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InICLR, 2014

    2014

  32. [32]

    Understanding diffusion objectives as the ELBO with simple data augmentation

    Diederik P Kingma and Ruiqi Gao. Understanding diffusion objectives as the ELBO with simple data augmentation. InNeurIPS, 2023. URL https://openreview.net/forum?id= NnMEadcdyD

    2023

  33. [33]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Applying guidance in a limited interval improves sample and distribution quality in diffusion models

    Tuomas Kynkäänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. InNeurIPS, 2024. 11

    2024

  35. [35]

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

    Jiachen Lei, Keli Liu, Julius Berner, Y HoiM, Hongkai Zheng, Jiahong Wu, and Xiangxiang Chu. There is no V AE: End-to-end pixel-space generative modeling via self-supervised pre-training. InICLR, 2026. URLhttps://openreview.net/forum?id=HbUoKPIZmp

    2026

  36. [36]

    Back to basics: Let denoising generative models denoise

    Tianhong Li and Kaiming He. Back to basics: Let denoising generative models denoise. In CVPR, 2026

    2026

  37. [37]

    Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Jihong Zhang, Chao Zhang, Meng Chen, Jie Liu, Zheng Fang, Weiyan Wang, Jinbao Xue,...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  38. [38]

    SDXL-Lightning: Progressive Adversarial Diffusion Distillation

    Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxl-lightning: Progressive adversarial diffusion distillation.arXiv preprint arXiv:2402.13929, 2024. URL https://arxiv.org/abs/2402. 13929

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [39]

    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. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors,ECCV, pages 740–755, Cham,

  40. [40]

    ISBN 978-3-319-10602-1

    Springer International Publishing. ISBN 978-3-319-10602-1

  41. [41]

    Evaluating text-to-visual generation with image-to-text generation

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In ECCV, 2024

    2024

  42. [42]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InICLR, 2023. URL https://openreview.net/forum? id=PqvMRDCJT9t

    2023

  43. [43]

    Rectified Flow: A Marginal Preserving Approach to Optimal Transport

    Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint arXiv:2209.14577, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  44. [44]

    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. InICLR, 2023. URL https://openreview.net/forum? id=XVjTT1nw5z

    2023

  45. [45]

    Albergo, Nicholas M

    Nanye Ma, Mark Goldstein, Michael S. Albergo, Nicholas M. Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. InECCV, 2024

    2024

  46. [46]

    Hpsv3: Towards wide-spectrum human preference score

    Yuhang Ma, Xiaoshi Wu, Keqiang Sun, and Hongsheng Li. Hpsv3: Towards wide-spectrum human preference score. InICCV, 2025

    2025

  47. [47]

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

    Zehong Ma, Longhui Wei, Shuai Wang, Shiliang Zhang, and Qi Tian. Deco: Frequency- decoupled pixel diffusion for end-to-end image generation. InCVPR, 2026

    2026

  48. [48]

    PixelGen: Improving Pixel Diffusion with Perceptual Supervision

    Zehong Ma, Ruihan Xu, and Shiliang Zhang. Pixelgen: Pixel diffusion beats latent diffusion with perceptual loss.arXiv preprint arXiv:2602.02493, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  49. [49]

    A perceptual color space for image processing, 2020

    Björn Ottosson. A perceptual color space for image processing, 2020. URL https:// bottosson.github.io/posts/oklab/

    2020

  50. [50]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InICCV, 2023

    2023

  51. [51]

    SDXL: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. InICLR, 2024. URL https://openreview.net/forum?id=di52zR8xgf. 12

    2024

  52. [52]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InICML, pages 8748–8763, 2021

    2021

  53. [53]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. InCVPR, 2022

    2022

  54. [54]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Inter- vention (MICCAI), pages 234–241, 2015

    2015

  55. [55]

    Seyedmorteza Sadat, Otmar Hilliges, and Romann M. Weber. Eliminating oversaturation and artifacts of high guidance scales in diffusion models. InICLR, 2025

    2025

  56. [56]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. InNeurIPS, 2022

    2022

  57. [57]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. InICLR, 2022

    2022

  58. [58]

    LAION-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wight- man, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text m...

    2022

  59. [59]

    Schönemann

    Peter H. Schönemann. A generalized solution of the orthogonal procrustes problem.Psychome- trika, 31(1):1–10, 1966. doi: 10.1007/BF02289451

    work page doi:10.1007/bf02289451 1966

  60. [60]

    Representation alignment for just image transformers is not easier than you think.arXiv preprint arXiv:2603.14366, 2026

    Jaeyo Shin, Jiwook Kim, and Hyunjung Shim. Representation alignment for just image transformers is not easier than you think.arXiv preprint arXiv:2603.14366, 2026

    work page arXiv 2026

  61. [61]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. InICML, pages 2256–2265, 2015

    2015

  62. [62]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. InNeurIPS, 2019

    2019

  63. [63]

    Score-based generative modeling through stochastic differential equations

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

    2021

  64. [64]

    arXiv preprint arXiv:2601.16208 , year=

    Shengbang Tong, Boyang Zheng, Ziteng Wang, Bingda Tang, Nanye Ma, Ellis Brown, Jihan Yang, Rob Fergus, Yann LeCun, and Saining Xie. Scaling text-to-image diffusion transformers with representation autoencoders.arXiv preprint arXiv:2601.16208, 2026

    work page arXiv 2026

  65. [65]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, Tianxing Wa...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  66. [66]

    Pixnerd: Pixel neural field diffusion

    Shuai Wang, Ziteng Gao, Chenhui Zhu, Weilin Huang, and Limin Wang. Pixnerd: Pixel neural field diffusion. InICLR, 2026. URLhttps://openreview.net/forum?id=BDnOrExHmt. 13

    2026

  67. [67]

    Ddt: Decoupled diffusion transformer

    Shuai Wang, Zhi Tian, Weilin Huang, and Limin Wang. Ddt: Decoupled diffusion transformer. InCVPR, 2026

    2026

  68. [69]

    URLhttps://arxiv.org/abs/2508.02324

    work page internal anchor Pith review Pith/arXiv arXiv

  69. [70]

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

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341, 2023. URL https://arxiv.org/ abs/2306.09341

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  70. [71]

    Jaakkola

    Yilun Xu, Shangyuan Tong, and Tommi S. Jaakkola. Stable target field for reduced variance score estimation in diffusion models. InICLR, 2023. URL https://openreview.net/ forum?id=WmIwYTd0YTF

    2023

  71. [72]

    Reconstruction vs

    Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimiza- tion dilemma in latent diffusion models. InCVPR, 2025

    2025

  72. [73]

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

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. InICLR, 2025

    2025

  73. [74]

    Pixeldit: Pixel diffusion transformers for image generation

    Yongsheng Yu, Wei Xiong, Weili Nie, Yichen Sheng, Shiqiu Liu, and Jiebo Luo. Pixeldit: Pixel diffusion transformers for image generation. InCVPR, 2026

    2026

  74. [75]

    Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer

    Z-Image Team, Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Xin Jin, Liangchen Li, Zhen Li, Zhong-Yu Li, David Liu, Dongyang Liu, Junhan Shi, Qilong Wu, Feng Yu, Chi Zhang, Shifeng Zhang, and Shilin Zhou. Z-image: An efficient image generation foundation model with single-stream diffusion transformer.ar...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  75. [76]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InCVPR, 2018

    2018

  76. [77]

    Unipc: A unified predictor- corrector framework for fast sampling of diffusion models

    Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor- corrector framework for fast sampling of diffusion models. InNeurIPS, 2023

    2023

  77. [78]

    starwars

    Boyang Zheng, Nanye Ma, Shengbang Tong, and Saining Xie. Diffusion transformers with representation autoencoders. InICLR, 2026. URL https://openreview.net/forum?id= 0u1LigJaab. 14 A Method Details A.1 Low-Rank Subspace Construction For transformer-based pixel generation, AsymFlow requires a patch-wise low-rank subspace. We use two constructions, depending...

    2026