Pith. sign in

REVIEW 3 major objections 5 minor 10 cited by

Transition Models: Rethinking the Generative Learning Objective

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A new training objective lets one model generate images at any step count, improving as steps increase.

desk verdict Strong empirical advance, overstated theory: Eq. (8) is a self-consistency condition, not an exact identity, and the bootstrapped objective needs a convergence argument. read the letter →

arxiv 2509.04394 v1 pith:SI2PQHNJ submitted 2025-09-04 cs.LG cs.CV

Zidong Wang , Yiyuan Zhang , Xiaoyu Yue , Xiangyu Yue , Yangguang Li , Wanli Ouyang , Lei Bai This is my paper
classification cs.LGcs.CV
keywords transitionmodelsstateidentityarbitrary-stepsamplingdiffusionconsistencytext-to-imagegenerationmonotonicrefinementlearningobjective
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to resolve the usual speed–quality trade-off in generative image modeling by changing the training objective rather than the architecture or the sampler. It argues that a single objective—supervising exact state-to-state transitions over arbitrary time intervals—should let one network act as a one-step generator, a few-step sampler, and a many-step refined integrator at the same time. The claim is that this removes the quality ceiling of few-step models and the step-count cost of diffusion models: quality should rise monotonically as function evaluations are added. Empirically, an 865M-parameter TiM trained from scratch reports GenEval scores of 0.67 at 1 NFE and 0.83 at 128 NFE, above the 0.70 and 0.65 reported for 8B and 12B baselines at every evaluated step count.

What carries the argument

The load-bearing object is the State Transition Identity, a product-derivative invariant d(B_{t,r}h(t))/dt = 0 with h(t) = α̂_t x + σ̂_t ε − f_{θ,t,r}. It forces two things at once: implicit trajectory consistency (a direct jump equals a composition of jumps) and time-slope matching (the model is penalized for the temporal change of its own residual). Training uses the Differential Derivation Equation (DDE), a finite-difference rule for the network's time derivative, which keeps the objective forward-only and compatible with fused attention and sharded data-parallel training. The architecture also adds decoupled absolute-time and interval embeddings plus interval-aware attention so the model

What would settle it

Pick a trained TiM and many random (x, ε, t, r). Compute B_{t,r}(α̂_t x + σ̂_t ε − fθ(xt,t,r)) at two nearby starting times, divide the difference by the time gap, and check whether Eq. (8)'s left-hand side is near zero; if it drifts significantly for large intervals, the identity is a condition the model fails, not an exact law. A second test: from fixed noise seeds, run the same checkpoint at 1, 8, 32, and 128 steps and check monotonicity of GenEval/FID—any non-monotonic pair would contradict the central scalability claim.

Watch

Extended reading notes

Core claim

The central claim is the State Transition Identity, Eq. (8): the time derivative of the product B_{t,r}(α̂_t x + σ̂_t ε − f_{θ,t,r}) is zero. Here B_{t,r} is a known coefficient from the noise schedule and f_{θ,t,r} is the network's prediction conditioned on both the current time t and the target time r. The identity says a direct transition from t to r must equal any composition of shorter transitions, and that the residual's value and its time slope are constrained together. From it the paper derives a training target Eq. (9) that generalizes ordinary diffusion supervision in the small-interval limit and consistency-model supervision at r=0 as special cases. The paper's claim is that a net

Load-bearing premise

TiM assumes that a self-referential target—computed from the network's own time derivative via a finite-difference approximation—converges to the true state-to-state transition over every interval, and that this derivative estimate stays accurate enough for large jumps and high resolutions.

Editorial extensions

If this is right

  • A single checkpoint trained from scratch can serve 1-step, few-step, and many-step sampling, eliminating the need for separate distilled models or teacher networks.
  • Quality should improve monotonically with NFEs, so users can trade compute for quality on the same model without switching checkpoints.
  • The objective subsumes diffusion training and consistency training as limits, giving a unified view of local versus finite-horizon supervision.
  • The forward-only derivative estimate removes a main scaling blocker of JVP-based consistency training, making the method practical for billion-parameter text-to-image pretraining.
  • The same identity applies across standard transports and supports native-resolution generation up to 4096×4096.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The identity's semigroup property suggests TiM should also be robust to arbitrary, even adaptive, step schedules; the paper evaluates fixed schedules only, so adaptive step selection is a natural testable extension.
  • Because the target is bootstrapped from the network's own derivative, a convergence check on held-out trajectories—does the learned map actually satisfy Eq. (8)?—would be a sharper validation than FID alone.
  • The r=0 and t=r limits imply one could dial the mix of diffusion-style and consistency-style supervision; the paper fixes that mix empirically, and tuning it per dataset could change the few-step versus many-step balance.
  • If monotone improvement holds beyond GenEval and ImageNet, sampling compute becomes a clean quality knob for deployment, with implications for scheduling in serving systems.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces Transition Models (TiM), a diffusion-style generative model trained to predict state-to-state transitions over arbitrary intervals Δt instead of only local PF-ODE velocities or fixed endpoint maps. The authors derive a "State Transition Identity" (Eq. 8) and use it to construct a training target that includes the network's own time derivative, computed via a finite-difference "DDE" to avoid JVP-based backpropagation. They also propose decoupled time/interval embeddings and interval-aware attention. On ImageNet-256, the training objective lowers 1-NFE FID from 309.5 to 49.91, and the final architecture/weighting reaches 47.46/24.62/17.10 at 1/8/50 NFEs. On text-to-image, an 865M-parameter model trained from scratch on 33M images achieves GenEval scores of 0.67 at 1 NFE and 0.83 at 128 NFE, outperforming SD3.5-Large (8B) and FLUX.1-Dev (12B) at the same step counts, with additional results on MJHQ30K, DPGBench, high resolutions, and multiple aspect ratios.

Significance. If the theoretical framing were secure, the contribution would be major: a single from-scratch model that spans one-step and many-step generation with monotonic quality improvement, and a JVP-free derivative computation that makes the objective scalable to billion-parameter training. The empirical work is rich and generally well executed: Table 4 ablates objective, architecture, and training choices; Tables 9–12 systematically compare transports, DDE epsilon values, timestep mixtures, and time weighting functions; and the code is available. The central weakness is that the claimed "exact identity" is in fact a self-consistency condition, and the fixed-point objective (Eq. 9) has no convergence proof. The empirical results are suggestive, but the paper's central theoretical claim needs substantial revision before the contribution can be accepted as stated.

major comments (3)
  1. [§3.2, Eqs. (6)–(8)] The "State Transition Identity" is not an identity for an arbitrary network. Eq. (7) is obtained by differentiating Eq. (6) and setting dx_r/dt = 0; this presupposes that Eq. (6) already holds exactly for every t along the trajectory. For a general f_θ, Eq. (6) defines a predicted x_r that is t-dependent, so differentiating it yields a consistency condition that the model must satisfy, not a property of the PF-ODE. In fact, d(B_{t,r}h(t))/dt = 0 has solutions B_{t,r}h(t) = C; the choice C = 0 is an additional boundary condition, not derived. The paper should either prove that the fixed point of Eq. (13) implies Eq. (8) with h = 0, or reframe Eq. (8) as a training objective/regularizer rather than an exact identity.
  2. [§3.3, Eqs. (9)–(10)] The training target in Eq. (9) is a fixed-point objective: it uses the network's own time derivative df_{θ−}/dt. Eq. (13) therefore minimizes the distance to a target that depends on the current model, and the paper provides no convergence guarantee for this self-distillation loop. Moreover, the general solution of d(Bh)/dt = 0 includes h = C/B, so Eq. (9) alone does not identify the true transition operator without an additional boundary condition. The DDE finite-difference approximation (Eq. 10) at ε = 0.005 is validated only through FID on ImageNet-256 (Table 10); there is no error bound and no validation for large intervals or high resolutions such as 4096×4096, where the derivative term is load-bearing.
  3. [§4.3, Table 4 and Appx. D.2, Table 11] The reported gains over the SiT baseline cannot be fully attributed to the proposed transition identity. The final training recipe includes 50% t = r samples (standard diffusion supervision) and 10% r = 0 samples (consistency-style supervision), as shown in Table 11, and also uses model-guidance (Eq. 51). Table 4 compares the TiM objective against a plain SiT baseline, but does not include a control that mixes standard diffusion and consistency objectives without the fixed-point TiM target. Such a control is needed to establish that the improvement stems from Eq. (8)–(13) rather than from the additional supervision mixture.
minor comments (5)
  1. [Appx. A.3, Algorithm 2] The sampling algorithm's indexing appears inconsistent: with t_N = T and t_0 = 0, the loop should update x_{t_{i-1}} from x_{t_i}, but the pseudocode writes x_{t_{i+1}} in both places. The stochastic-correction formulas also appear malformed (missing operators in the ε̂ expression).
  2. [Eq. (51)] The model-guidance target introduces f^cond and f^uncond without defining the interval arguments; it also uses the EMA weights θ* but does not explain how the DDE derivative is computed with respect to the EMA model. This makes the final training target ambiguous.
  3. [Table 4] The grouping of ablations is confusing: rows (a)–(c) are labeled "training objective," while rows (d)–(g) are architecture ablations, but it is unclear which architecture is used for rows (b) and (c). Specify this explicitly.
  4. [Table 7, Eqs. (37)–(38)] The EDM TiM parameterization contains apparent typesetting errors in the denominators (e.g., "t2 + σ3_d" and "t2 + σ^3_d"), making the formulas difficult to verify. Please correct.
  5. [§3.2, after Eq. (8)] The domain of validity of the identity is not stated. B_{t,r} and dB_{t,r}/dt can vanish (e.g., at t = r), so the training target in Eq. (9) is singular in those limits; the paper relies on the mixture in Table 11 to avoid the issue but should state this explicitly.

Circularity Check

2 steps flagged · score 4.0 of 10

State Transition Identity is a self-consistency condition and the training target bootstraps the network's own derivative; external benchmarks keep this from being fully circular.

  1. self definitional [Section 3.2, Eq. (6) and Eq. (8)]
    "The derivation begins with the general mathematical form for a state transition between points (xt, xr) on a PF-ODE trajectory, as given in Eq. (6). The central principle is to treat this form not as a numerical approximation, but as an exact identity that must hold for any interval ∆t = t − r. ... We denote Equation (8) as the State Transition Identity, a product-derivative invariant."

    Eq. (6) only defines a predicted state x̂_r = A_{t,r}x_t + B_{t,r}f_θ; it equals the true x_r only when f_θ is the exact transition operator. Differentiating Eq. (6) and setting dx_r/dt = 0 is exactly the condition that this predicted state is constant, i.e. d/dt(B_{t,r}(α̂_t x + σ̂_t ε − f_θ)) = 0. Thus Eq. (8) restates the self-consistency condition that defines what f_θ must satisfy. Presenting it as a derived 'exact identity' assumes the very property the model is supposed to learn, so the claimed first-principles result reduces by construction to the model's defining condition.

  2. fitted input called prediction [Section 3.2, Eq. (9); Section 3.3, Eq. (13)]
    "Derived from State Transition Identity (Eq. (8)), we obtain the learning target ˆf = ˆαtx + ˆσtε + Bt,r / dBt,r dt ( d ˆαt dt x + dˆσt dt ε − dfθ−,t,r dt ), where θ− indicates the fixed network parameter θ and dfθ−,t,r dt is the time derivative of the network."

    The regression target for fθ is computed from fθ− and its time derivative, so the 'ground-truth' transition for every interval (t,r) is generated by the current model itself. This is a fixed-point/self-distillation objective, not supervision from an independent exact transition. The paper provides no convergence proof for this bootstrap, and the DDE finite-difference (Eq. 10) is an uncontrolled approximation, so the claim that Eq. (13) 'analytically defines state transitions' and yields monotonic improvement depends on an unproved fixed-point convergence.

full rationale

The central theoretical derivation is partially circular: Eq. (8) is not an exact identity for an arbitrary network but a consistency condition that the model must satisfy to be an exact transition operator, and Eq. (9) defines the training target using the network's own derivative, making the objective self-referential. However, the paper is not wholly circular: the claimed empirical results are evaluated on external benchmarks (GenEval, MJHQ30K, ImageNet), the architecture ablations are independent, and the training objective is an implementable self-distillation scheme rather than a fitted parameter renamed as a prediction. The mathematical overstatement and missing convergence proof are serious rigor concerns, but they do not force the benchmark conclusions by construction. Hence a score of 4 is appropriate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim rests on several hand-picked hyperparameters (weighting, epsilon, timestep sampling, guidance scale) and on at least two unproven assumptions: convergence of the bootstrapped objective and accuracy of the DDE across scales. No new physical entities are introduced. The mathematical identity is a restatement of model exactness, not an independent constraint.

free parameters (5)
  • Time weighting w(t,r) = (σ_data + tan(t) − tan(r))^(−1/2)
    Hand-picked after ablating several transforms and kernels on ImageNet (Table 12). This weights short-interval transitions more heavily and is a free choice affecting training dynamics and final quality.
  • DDE epsilon = 0.005
    Chosen from a sweep (Table 10) using the small TiM-B/4 model; assumed to transfer to the 865M T2I model without retuning.
  • Timestep sampling mixture = 50% t=r, 10% r=0, 40% random intervals
    Ablated on ImageNet (Table 11). This mixture balances pure diffusion supervision and consistency-style supervision; it is a design choice, not derived from the theory.
  • Model-guidance scale ω = 1.75
    Applied after 100K iterations for T2I training; adopted from Tang et al. [73] with no ablation reported for TiM.
  • LoRA rank in AdaLN = r = D/3
    Architecture hyperparameter for the decoupled time/interval embedding; chosen to limit parameter overhead.
assumptions (4)
  • domain assumption The PF-ODE and its solution manifold are well-defined and the network fθ can represent transitions between any two states on the trajectory.
    The whole framework assumes a smooth generative trajectory and that a neural network with decoupled time/interval embeddings can approximate the solution manifold. Invoked in Sec. 3.2 and the architecture section.
  • ad hoc to paper The self-consistent fixed-point training objective Eq. (9), which includes the network's own time derivative, converges to the exact transition operator.
    This is not proven. The paper validates it empirically only through small-scale ImageNet ablations, and no convergence argument is given. Used to justify the training target and the monotonic improvement claim.
  • ad hoc to paper The finite-difference approximation in Eq. (10) is accurate enough to replace the Jacobian-vector product, including for large intervals and high-resolution inputs.
    The DDE is a central enabler of scalability, but its accuracy is only shown for FID on ImageNet across epsilon values. Its validity at 4096x4096 resolution and with the model-guidance target is assumed.
  • standard math Standard diffusion theory: the forward SDE, the reverse-time SDE, and the PF-ODE equivalence from Song et al. (2020).
    Used throughout Sec. 3.1 and Appx. A to define the data-to-noise process and the transition coefficients.
invented entities (1)
  • State Transition Identity (Eq. 8)
    purpose: A product-derivative invariant used to derive the training target for arbitrary-interval transitions.
    It is a reformulation of the condition that the model's transition is exact, not an independently falsifiable object. Its predictive content is entirely carried by the trained network.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transition Models: Rethinking the Generative Learning Objective." pith.science (2026). https://pith.science/paper/SI2PQHNJ

@misc{pith2026250904394,
  author       = {Pith},
  title        = {Pith review of: Transition Models: Rethinking the Generative Learning Objective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SI2PQHNJ}},
  note         = {Machine review of arXiv:2509.04394}
}
read the original abstract

A fundamental dilemma in generative modeling persists: iterative diffusion models achieve outstanding fidelity, but at a significant computational cost, while efficient few-step alternatives are constrained by a hard quality ceiling. This conflict between generation steps and output quality arises from restrictive training objectives that focus exclusively on either infinitesimal dynamics (PF-ODEs) or direct endpoint prediction. We address this challenge by introducing an exact, continuous-time dynamics equation that analytically defines state transitions across any finite time interval. This leads to a novel generative paradigm, Transition Models (TiM), which adapt to arbitrary-step transitions, seamlessly traversing the generative trajectory from single leaps to fine-grained refinement with more steps. Despite having only 865M parameters, TiM achieves state-of-the-art performance, surpassing leading models such as SD3.5 (8B parameters) and FLUX.1 (12B parameters) across all evaluated step counts. Importantly, unlike previous few-step generators, TiM demonstrates monotonic quality improvement as the sampling budget increases. Additionally, when employing our native-resolution strategy, TiM delivers exceptional fidelity at resolutions up to 4096x4096.

Figures

Figures reproduced from arXiv: 2509.04394 by the authors.

Figure 1
Figure 1. TiM’s superior performance across different NFEs, resolutions, and aspect ratios. On the GenEval [27] bench￾mark, TiM outperforms Flux.1 models [5, 6] at different NFEs (top, 1024 × 1024), at higher resolutions (middle, 1024 × 1024 to 4096 × 4096), and diverse aspect ratios (bottom, 2 : 5 to 5 : 2). Function Evaluations (NFEs)—approximately proportional to the number of integration steps—leading to increased in￾fere… view at source ↗
Figure 2
Figure 2. Illustration of Different Generative Paradigms. While conventional diffusion models learn the local vector field and few-step models learn a fixed endpoint map (a single large step), our Transition Models (TiM) are trained to master arbitrary state-to-state transitions. This approach allows TiM to learn the entire solution manifold of the generative process, unifying the few-step and many-step regimes within a singl… view at source ↗
Figure 3
Figure 3. Qualitative Analysis between TiM and existing methods under different NFEs. TiM delivers superior fidelity and text alignment across all NFEs. In contrast, multi-step diffusion and few-step distilled models exhibit pronounced step–quality trade-offs: SDXL, SD3.5-Large, and FLUX.1-Dev fail to generate images at low NFEs, while SDXL-Turbo, SD3.5-Turbo, and FLUX.1-Schnell produce over-saturated outputs at high NFEs. of… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: TiM Model Architecture. A magpie with a glossy black head, wings, and tail, and a white belly, perches on a wooden post of a wire fence. Self-Attn Block Feed Forward Cross-Attn Block LoRA AdaLN LLM Decoder �1 , �1 �1 �2 , �2 �2 �3 , �3 �3 Patcify Timestep & Interval × …
Figure 6
Figure 6. Figure 6: High-resolution and multi-aspect generations from TiM (128 NFEs). TiM attains up to 4096 × 4096 resolution and reliably handles multiple aspect ratios, including 1024 × 4096 and 2560 × 1024. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 10 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation

    cs.CV 2026-05 unverdicted novelty 8.0 of 10

    AnyFlow enables any-step video diffusion by distilling flow-map transitions over arbitrary time intervals with on-policy backward simulation.

  2. One-Step Generative Modeling via Wasserstein Gradient Flows

    cs.LG 2026-05 conditional novelty 7.0 of 10

    W-Flow achieves state-of-the-art one-step ImageNet 256x256 generation at 1.29 FID by training a static neural network to follow a Wasserstein gradient flow that minimizes Sinkhorn divergence, delivering roughly 100x f...

  3. Parallel Decoding Distillation for Fast Image and Video Generation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A trajectory-based distillation method trains a student to predict multiple mean velocities per network evaluation, enabling 4-8 step generation with competitive quality and improved diversity.

  4. One-Step Generative Modeling via Wasserstein Gradient Flows

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    W-Flow compresses a Wasserstein gradient flow defined via Sinkhorn divergence into a single-step neural generator, reporting 1.29 FID on ImageNet 256x256 with improved mode coverage.

  5. Extending One-Step Image Generation from Class Labels to Text via Discriminative Text Representation

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    By requiring and using highly discriminative LLM text features, the work enables the first effective one-step text-conditioned image generation with MeanFlow.

  6. Transition Matching Distillation for Fast Video Generation

    cs.CV 2026-01 conditional novelty 6.0 of 10

    Splitting a video diffusion model into a fixed feature extractor and a small recurrent flow head lets TMD generate videos in one to two effective steps with better VBench scores than prior distilled models.

  7. Understanding, Accelerating, and Improving MeanFlow Training

    cs.CV 2025-11 conditional novelty 6.0 of 10

    Training MeanFlow by first forming instantaneous velocity and short-gap average velocity, then shifting to long gaps, improves 1-NFE ImageNet FID from 3.43 to 2.87 and speeds training by about 2.5x.

  8. Trajectory Forcing: Structure-First Generation with Controllable Semantic Trajectories

    cs.CV 2026-06 unverdicted novelty 5.0 of 10

    Trajectory Forcing makes generative image synthesis trajectory-centric by organizing it into decodable semantic stages derived from clustered visual representations and trained with one-step flow-matching models.

  9. SubFlow: Sub-mode Conditioned Flow Matching for Diverse One-Step Generation

    cs.LG 2026-04 unverdicted novelty 5.0 of 10

    SubFlow restores full mode coverage in one-step flow matching by conditioning on sub-modes from semantic clustering, yielding higher diversity on ImageNet-256 while preserving FID.

  10. Improved Mean Flows: On the Challenges of Fastforward Generative Models

    cs.CV 2025-12 unverdicted novelty 5.0 of 10

    Improved MeanFlow (iMF) reaches 1.72 FID on ImageNet 256x256 with one function evaluation by reformulating the training objective as a regression on instantaneous velocity and treating guidance as flexible conditionin...

Reference graph

Works this paper leans on

93 extracted references · 34 canonical work pages · cited by 9 Pith papers

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

    arXiv

  2. [2]

    Megalith-huggingface

    adebyollin. Megalith-huggingface. https://huggingface.co/datasets/madebyollin/megalith-10m. 6

  3. [3]

    Reverse-time diffusion equation mod- els

    Brian DO Anderson. Reverse-time diffusion equation mod- els. Stochastic Processes and their Applications, 12(3):313– 326, 1982. 3

    1982

  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. In CVPR, 2023. 23

    2023

  5. [5]

    Flux.1-dev

    black-forest labs. Flux.1-dev. https://huggingface. co/black-forest-labs/FLUX.1-dev , . 1, 2, 6, 7, 8

  6. [6]

    Flux.1-schnell

    black-forest labs. Flux.1-schnell. https : / / huggingface . co / black - forest - labs / FLUX.1-schnell, . 1, 6, 7, 8

  7. [7]

    Flow map matching with stochastic inter- polants: A mathematical framework for consistency models

    Nicholas Matthew Boffi, Michael Samuel Albergo, and Eric Vanden-Eijnden. Flow map matching with stochastic inter- polants: A mathematical framework for consistency models. Transactions on Machine Learning Research, 2025. 1, 2

    2025

  8. [8]

    Large scale gan training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 22, 23

    arXiv 2018

Show all 93 references
  1. [9]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luh- man, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators

  2. [10]

    Maskgit: Masked generative image transformer

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 22

    2022

  3. [11]

    Conceptual 12m: Pushing web-scale image-text pre- training to recognize long-tail visual concepts

    Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre- training to recognize long-tail visual concepts. In CVPR,

  4. [12]

    Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6

    2023 arXiv

  5. [13]

    Deep compression autoencoder for efficient high-resolution diffu- sion models

    Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffu- sion models. arXiv preprint arXiv:2410.10733, 2024. 6, 19, 21

    2024 arXiv

  6. [14]

    Pixart-sigma: Weak-to-strong train- ing 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-sigma: Weak-to-strong train- ing of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91. ...

    2024

  7. [15]

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset

    Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Sil- vio Savarese, et al. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset. arXiv preprint arXiv:2505.09568, 2025. 6

    2025 arXiv

  8. [16]

    Sana-sprint: One-step diffusion with continuous-time con- sistency distillation

    Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Song Han, and Enze Xie. Sana-sprint: One-step diffusion with continuous-time con- sistency distillation. arXiv preprint arXiv:2503.09641, 2025. 6

    2025

  9. [17]

    Flashattention-2: Faster attention with bet- ter parallelism and work partitioning

    Tri Dao. Flashattention-2: Faster attention with bet- ter parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 4

    2023 arXiv

  10. [18]

    Redcaps: Web-curated image-text data created by the people, for the people

    Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin John- son. Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431 ,

    arXiv

  11. [19]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 21, 23

    2021

  12. [20]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. 2024. 1, 2, 6, 7, 8, 20

    2024

  13. [21]

    One step diffusion via shortcut models

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. arXiv preprint arXiv:2410.12557, 2024. 1, 17, 19, 22

    2024 arXiv

  14. [22]

    Lumina-t2x: Transforming text into any modality, reso- lution, and duration via flow-based large diffusion transform- ers

    Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, reso- lution, and duration via flow-based large diffusion transform- ers. arXiv preprint arXiv:2405.05945, 2024. 22, 23

    2024 arXiv

  15. [24]

    Mdtv2: Masked diffusion transformer is a strong image synthesizer

    Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389,

    arXiv

  16. [25]

    Seedream 3.0 technical report

    Yu Gao, Lixue Gong, Qiushan Guo, Xiaoxia Hou, Zhichao Lai, Fanshi Li, Liang Li, Xiaochen Lian, Chao Liao, Liyang Liu, et al. Seedream 3.0 technical report. arXiv preprint arXiv:2504.11346, 2025. 7

    2025 arXiv

  17. [26]

    Mean flows for one-step genera- tive modeling

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step genera- tive modeling. arXiv preprint arXiv:2505.13447, 2025. 1, 2, 4, 17, 19, 22

    2025 arXiv

  18. [27]

    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. Advances in Neural Information Pro- cessing Systems, 36:52132–52152, 2023. 1, 2

    2023

  19. [28]

    Seedream 2.0: A native chinese-english bilin- gual image generation foundation model

    Lixue Gong, Xiaoxia Hou, Fanshi Li, Liang Li, Xiaochen Lian, Fei Liu, Liyang Liu, Wei Liu, Wei Lu, Yichun Shi, et al. Seedream 2.0: A native chinese-english bilin- gual image generation foundation model. arXiv preprint arXiv:2503.07703, 2025. 6, 7 9

    2025 arXiv

  20. [29]

    Diffit: Diffusion vision transformers for im- age generation

    Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, and Arash Vahdat. Diffit: Diffusion vision transformers for im- age generation. arXiv preprint arXiv:2312.02139, 2023. 23

    2023 arXiv

  21. [30]

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

    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 equilib- rium. NeurIPS, 2017. 21

    2017

  22. [31]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. NeurIPS, 2020. 2, 16, 20

    2020

  23. [32]

    sim- ple diffusion: End-to-end diffusion for high resolution im- ages

    Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. sim- ple diffusion: End-to-end diffusion for high resolution im- ages. 2023. 22, 23

    2023

  24. [33]

    text-to-image-2m

    jackyhate. text-to-image-2m. https://huggingface.co/datasets/jackyhate/text-to-image- 2M. 6

  25. [34]

    Scal- ing up gans for text-to-image synthesis

    Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scal- ing up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10124–10134, 2023. 22

    2023

  26. [35]

    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. NeurIPS, 2022. 1, 2, 3, 15, 16, 17, 20

    2022

  27. [36]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174–24184, 2024. 16, 17, 20, 23

    2024

  28. [37]

    Consistency trajectory mod- els: Learning probability flow ode trajectory of diffusion

    Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Mu- rata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory mod- els: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 17, 19

    2023 arXiv

  29. [38]

    Kynk ¨a¨anniemi, T

    T. Kynk ¨a¨anniemi, T. Karras, S. Laine, and T Lehtinen, J.and Aila. Improved precision and recall metric for assess- ing generative models. NeurIPS, 2019. 21

    2019

  30. [39]

    Playground v2

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Lin- miao Xu, and Suhail Doshi. Playground v2. 5: Three in- sights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 6

    2024 arXiv

  31. [40]

    Autoregressive image generation without vec- tor quantization

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vec- tor quantization. Advances in Neural Information Processing Systems, 37:56424–56445, 2024. 22

    2024

  32. [41]

    Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chi- nese understanding

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chi- nese understanding. arXiv preprint arXiv:2405.087...

    2024 arXiv

  33. [42]

    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. 1

    2024 arXiv

  34. [43]

    Flow matching for generative mod- eling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling. arXiv preprint arXiv:2210.02747, 2022. 3, 13, 16, 19, 20

    2022 arXiv

  35. [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. In ICLR, 2023. 2, 16, 20

    2023

  36. [45]

    Simplifying, stabilizing and scaling continuous-time consistency models

    Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1, 2, 3, 4, 13, 15, 16, 17, 20

    2024 arXiv

  37. [46]

    Dpm-solver: A fast ode solver for diffu- sion probabilistic model sampling in around 10 steps

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffu- sion probabilistic model sampling in around 10 steps. Ad- vances in neural information processing systems , 35:5775– 5787, 2022. 1, 2

    2022

  38. [47]

    Dpm-solver++: Fast solver for guided sam- pling of diffusion probabilistic models.Machine Intelligence Research, pages 1–22, 2025

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sam- pling of diffusion probabilistic models.Machine Intelligence Research, pages 1–22, 2025. 2, 4

    2025

  39. [48]

    Latent consistency models: Synthesizing high- resolution images with few-step inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high- resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 6

    2023 arXiv

  40. [49]

    Sit: Explor- ing flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Explor- ing flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740,

    arXiv

  41. [50]

    C. Nash, J. Menick, S. Dieleman, and P. W Battaglia. Gen- erating images with sparse representations. arXiv preprint arXiv:2103.03841, 2021. 21

    2021 arXiv

  42. [51]

    Randar: Decoder-only autoregressive visual generation in random orders

    Ziqi Pang, Tianyuan Zhang, Fujun Luan, Yunze Man, Hao Tan, Kai Zhang, William T Freeman, and Yu-Xiong Wang. Randar: Decoder-only autoregressive visual generation in random orders. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 45–55, 2025. 22

    2025

  43. [52]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1, 5, 6, 22, 23

    2023

  44. [53]

    Flow-anchored consistency mod- els

    Yansong Peng, Kai Zhu, Yu Liu, Pingyu Wu, Hebei Li, Xi- aoyan Sun, and Feng Wu. Flow-anchored consistency mod- els. arXiv preprint arXiv:2507.03738, 2025. 1, 2, 4, 17

    2025

  45. [54]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 6

    2023 arXiv

  46. [55]

    Hyper-sd: Trajec- tory segmented consistency model for efficient image syn- thesis

    Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajec- tory segmented consistency model for efficient image syn- thesis. Advances in Neural Information Processing Systems, 37:117340–117362, 2025. 2, 6

    2025

  47. [56]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 6, 21, 22

    2022

  48. [57]

    Align your flow: Scaling continuous-time flow map distillation

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your flow: Scaling continuous-time flow map distillation. arXiv preprint arXiv:2506.14603, 2025. 1, 2, 4

    2025 arXiv

  49. [58]

    Salimans, I

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Rad- ford, and X Chen. Improved techniques for training gans. NeurIPS, 2016. 21 10

    2016

  50. [59]

    Stylegan- xl: Scaling stylegan to large diverse datasets

    Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan- xl: Scaling stylegan to large diverse datasets. In ACM SIG- GRAPH 2022 conference proceedings, 2022. 23

    2022

  51. [60]

    Stylegan- xl: Scaling stylegan to large diverse datasets

    Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan- xl: Scaling stylegan to large diverse datasets. In ACM SIG- GRAPH 2022 conference proceedings, pages 1–10, 2022. 22

    2022

  52. [61]

    Fast high- resolution image synthesis with latent adversarial diffusion distillation

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high- resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia 2024 Conference Papers , pages 1–11, 2024. 1, 2, 6, 7, 8

    2024

  53. [62]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision , pages 87–103. Springer,

  54. [63]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

    Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 6

    2021 arXiv

  55. [64]

    Conceptual captions: A cleaned, hypernymed, im- age alt-text dataset for automatic image captioning

    Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, im- age alt-text dataset for automatic image captioning. 2018

    2018

  56. [65]

    From pixels to prose: A large dataset of dense image cap- tions

    Vasu Singla, Kaiyu Yue, Sukriti Paul, Reza Shirkavand, Mayuka Jayawardhana, Alireza Ganjdanesh, Heng Huang, Abhinav Bhatele, Gowthami Somepalli, and Tom Goldstein. From pixels to prose: A large dataset of dense image cap- tions. arXiv preprint arXiv:2406.10328, 2024. 6

    2024 arXiv

  57. [66]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 4

    2010 arXiv

  58. [67]

    Improved techniques for training consistency models

    Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In The Twelfth International Conference on Learning Representations. 1, 2, 17, 22

  59. [68]

    Generative modeling by es- timating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by es- timating gradients of the data distribution. NeurIPS, 2019. 16

    2019

  60. [69]

    Score-based generative modeling through stochastic differential equa- tions

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. arXiv preprint arXiv:2011.13456, 2020. 1, 2, 3, 13, 14, 15, 16, 20

    2011 arXiv

  61. [70]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, pages 32211–32252. PMLR, 2023. 1, 2, 17, 18

    2023

  62. [71]

    Autoregressive model beats diffusion: Llama for scalable image generation

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 22

    2024 arXiv

  63. [72]

    Unified continuous gener- ative models

    Peng Sun, Yi Jiang, and Tao Lin. Unified continuous gener- ative models. arXiv preprint arXiv:2505.07447, 2025. 3, 13, 15

    2025 arXiv

  64. [73]

    Diffusion models without classifier-free guidance

    Zhicong Tang, Jianmin Bao, Dong Chen, and Baining Guo. Diffusion models without classifier-free guidance. arXiv preprint arXiv:2502.12154, 2025. 20

    2025 arXiv

  65. [74]

    Gemma 3 technical report

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Ta- tiana Matejovicova, Alexandre Ram ´e, Morgane Rivi `ere, et al. Gemma 3 technical report. arXiv preprint arXiv:2503.19786, 2025. 6

    2025 arXiv

  66. [75]

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

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Li- wei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural in- formation processing systems , 37:84839–84865, 2024. 22, 23

    2024

  67. [76]

    Sta- ble consistency tuning: Understanding and improving con- sistency models

    Fu-Yun Wang, Zhengyang Geng, and Hongsheng Li. Sta- ble consistency tuning: Understanding and improving con- sistency models. arXiv preprint arXiv:2410.18958, 2024. 23

    2024 arXiv

  68. [77]

    Phased consistency models

    Fu-Yun Wang, Zhaoyang Huang, Alexander Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency models. Advances in Neural Information Pro- cessing Systems, 37:83951–84009, 2025. 2, 17, 19

    2025

  69. [78]

    Exploring dcn-like ar- chitecture for fast image generation with arbitrary resolu- tion

    Shuai Wang, Zexian Li, Tianhui Song, Xubin Li, Tiezheng Ge, Bo Zheng, and Limin Wang. Exploring dcn-like ar- chitecture for fast image generation with arbitrary resolu- tion. Advances in Neural Information Processing Systems , 37:87959–87977, 2024. 22, 23

    2024

  70. [79]

    Emu3: Next-token prediction is all you need

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 6

    2024 arXiv

  71. [80]

    Native-resolution image synthesis

    Zidong Wang, Lei Bai, Xiangyu Yue, Wanli Ouyang, and Yiyuan Zhang. Native-resolution image synthesis. arXiv preprint arXiv:2506.03131, 2025. 7

    2025 arXiv

  72. [81]

    Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer

    Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Chengyue Wu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, et al. Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. arXiv preprint arXiv:2501.18427, 2025. 6

    2025 arXiv

  73. [82]

    Exploring representation-aligned latent space for better generation

    Wanghan Xu, Xiaoyu Yue, Zidong Wang, Yao Teng, Wen- long Zhang, Xihui Liu, Luping Zhou, Wanli Ouyang, and Lei Bai. Exploring representation-aligned latent space for better generation. arXiv preprint arXiv:2502.00359 , 2025. 22

    2025 arXiv

  74. [83]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1

    2024 arXiv

  75. [84]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 6613–6623...

    2024

  76. [85]

    Improved distribution matching distillation for fast image synthesis

    Tianwei Yin, Micha ¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. Advances in Neural Information Processing Systems , 37: 47455–47487, 2025. 2

    2025

  77. [86]

    Language model beats diffusion–tokenizer is key to visual generation

    Lijun Yu, Jos ´e Lezama, Nitesh B Gundavarapu, Luca Ver- sari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh 11 Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion–tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 22, 23

    2023 arXiv

  78. [87]

    Representation alignment for generation: Training diffu- sion 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 diffu- sion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 22, 23

    2024 arXiv

  79. [88]

    Diffu- sion models need visual priors for image generation

    Xiaoyu Yue, Zidong Wang, Zeyu Lu, Shuyang Sun, Meng Wei, Wanli Ouyang, Lei Bai, and Luping Zhou. Diffu- sion models need visual priors for image generation. arXiv preprint arXiv:2410.08531, 2024. 22

    2024 arXiv

  80. [89]

    Pytorch fsdp: experi- ences on scaling fully sharded data parallel

    Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien- Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experi- ences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023. 4, 20

    2023 arXiv

  81. [90]

    Fast training of diffusion models with masked transformers

    Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. 2023. 23

    2023

  82. [91]

    Dpm- solver-v3: Improved diffusion ode solver with empirical model statistics

    Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Dpm- solver-v3: Improved diffusion ode solver with empirical model statistics. Advances in Neural Information Process- ing Systems, 36:55502–55542, 2023. 2, 4

    2023

  83. [92]

    Inductive moment matching

    Linqi Zhou, Stefano Ermon, and Jiaming Song. Inductive moment matching. arXiv preprint arXiv:2503.07565, 2025. 22

    2025 arXiv

  84. [93]

    Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation

    Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In Forty-first International Confer- ence on Machine Learning, 2024. 1 12 Appendix W...

    2024

  85. [2024]

    Accessed: 2024-5-1. 1

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.