Pith. sign in

REVIEW 4 major objections 4 minor 66 references

A Survey on Pre-Trained Diffusion Model Distillations

T0 review · 4 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read This survey organizes pre-trained diffusion-model distillation into three families under a single notation, so that existing methods can be compared directly.

desk verdict A timely and useful taxonomy of diffusion distillation, but the unified notation is undermined by misprinted equations and undefined terms; worth a serious referee only after substantial fixing. read the letter →

arxiv 2502.08364 v2 pith:DSCQJXP7 submitted 2025-02-12 cs.LG

classification cs.LG
keywords diffusionmodelsknowledgedistillationfidelitytrajectoryadversarialconsistencyrectifiedflowfew-stepgeneration
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

This paper is a survey whose claim is that the many ways to distill a pre-trained diffusion model can be systematically organized into three method families: fidelity distillation, trajectory distillation, and adversarial distillation. Fidelity methods match teacher outputs, whether image values, image distributions, one-step denoised images, or Fisher divergences; trajectory methods match or straighten the noise-to-data path; adversarial methods use a discriminator loss, either standalone or as an add-on. The paper's added value is a single notation system that rewrites each method's objective, motivation, and procedure so that methods from different families can be compared on equal terms. If this map is reliable, it gives practitioners and researchers a structured way to choose, combine, and extend distillation techniques for fast, low-resource generation.

What carries the argument

The central machinery is the three-way taxonomy itself, expressed through a shared notation: time steps live in $[0,1]$, $x_t$ denotes the noisy image at time $t$, $\epsilon_\theta(x_t,t)$ is the teacher score model, $v_\theta(x_t,t)$ is the teacher rectified-flow velocity, and $f_\phi$ is the student generator. Each method's loss is rewritten in this notation, which is what enables direct side-by-side comparison of methods as different as SDS and consistency distillation. The taxonomy's categories correspond to what the student is trained to match: output values, output distributions, denoised image spaces, Fisher divergences, trajectory self-consistency, trajectory straightness, or an adversarial discriminator's judgment.

What would settle it

Comparing Eq. (3) with the DDIM update in the cited source shows that the survey's version is missing parentheses that define which terms are divided by $\sqrt{\alpha_t}$, making the update ambiguous; a reader can also take the survey's VSD gradient formula, implement it for a one-step generator, and check it against automatic differentiation, where the stray symbol should make the gradient non-evaluable or numerically inconsistent.

Watch

Extended reading notes

Core claim

The paper's central claim is that pre-trained diffusion-model distillation is best understood by asking which teacher quantity the student is trained to reproduce. Fidelity distillation covers output reconstruction loss (denoising student, progressive distillation, TRACT), output distribution loss (SDS, VSD, Diff-Instruct, DMD), one-step denoising image space methods (SDI, Boot), and Fisher divergence loss (SiD and successors). Trajectory distillation covers consistency distillation (CM, CTM, continuous-time consistency models), rectified flow distillation (rectified flow, reflow, InstaFlow, SlimFlow, rectified flow++), and methods integrating consistency with straightness (Shortcut model, SCoT, Consistency-FM). Adversarial distillation covers Diffusion-GAN and adversarial diffusion distillation, and the paper argues adversarial loss is orthogonal and can be plugged into other distillation methods. The paper further claims that using one notation for all these methods makes their motivations, procedures, and objective functions conveniently comparable and precisely evaluable.

Load-bearing premise

The survey's comparative value rests on its claim that each cited method is faithfully restated in one notation, but the printed text already shows slips, including a missing parentheses in the DDIM update Eq. (3) and a stray symbol in the VSD gradient, so that fidelity of translation is not guaranteed.

Editorial extensions

If this is right

  • A practitioner can classify any new distillation method by asking which teacher quantity it matches: output value, output distribution, denoised image, Fisher divergence, trajectory self-consistency, trajectory straightness, or adversarial signal.
  • Because adversarial loss is orthogonal to the other families, it can be added to fidelity or trajectory methods, with SiDA and ADD given as examples.
  • The taxonomy separates step-count reduction from model-size reduction; the survey notes most methods focus on cutting steps and identifies smaller student models as an open direction.
  • The unified notation exposes structural coincidences across methods, such as SDS being a special case of VSD and CTM, Shortcut, and SCoT sharing a projection-based trajectory view.
  • The survey identifies missing standardized training guidelines and weight-function choices as reasons why distillation results across papers are hard to compare.

Reading between the lines

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

  • A reader could extend the taxonomy into a two-axis map, with loss family on one axis and trajectory-versus-output focus on the other, which would better accommodate hybrids like SiDA (adversarial plus Fisher divergence) and Shortcut/SCoT (consistency plus straightness).
  • The common notation suggests a testable benchmarking program: rewrite candidate distillation losses in this notation and compare them on identical student architectures and datasets, which the survey itself does not carry out.
  • The challenges section implies that distilling Stable Diffusion-class models into genuinely smaller-width students, not just few-step students, is a concrete direction that current methods mostly leave unmeasured.
  • The survey's omission of practical application details leaves room for a follow-up study on how these distillation families transfer to audio, video, and 3D generation, which the survey names as future work.
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

4 major / 4 minor

Summary. This survey reviews distillation methods for pre-trained diffusion models and organizes them into three categories: fidelity distillation, trajectory distillation, and adversarial distillation. For each category it provides a common notation, brief descriptions of representative methods, and a visual taxonomy in Figure 1. The paper also lists challenges and future directions, including smaller student models, trajectory optimization, and theoretical understanding. The central contribution is a unified methodological map that should allow direct comparison of distillation objectives across methods.

Significance. If the equations and descriptions are faithful, the survey would be a useful and timely resource: it covers recent developments such as consistency models, rectified flow distillation, and adversarial distillation, and it attempts a coherent taxonomy with shared notation. The paper includes a helpful taxonomy figure, a visual depiction of trajectory methods, and an up-to-date reference list. However, the survey's value rests on the accuracy of the equations it uses to characterize each method; as it stands, several anchor equations are incorrect, so the promised precision of the comparison framework is not yet delivered.

major comments (4)
  1. [Preliminary, Eq. (3)] The DDIM update is misprinted. The printed formula scales x_t by sqrt(alpha_{t-1}) only, whereas the correct DDIM step is x_{t-1} = sqrt(alpha_{t-1}) * (x_t - sqrt(1-alpha_t) eps_theta) / sqrt(alpha_t) + sqrt(1-alpha_{t-1}) eps_theta. As printed, the update changes the sampler's dynamics and misrepresents the teacher model's trajectory, which is load-bearing for the trajectory distillation section that builds on this equation.
  2. [Adversarial Loss, Eq. (7)] The ADD objective is not correctly stated. The expectation is over generator samples x'_t only, yet the first term log D_psi(x_t) uses an undefined x_t, and the real-data expectation term E_{x ~ p_real}[log D_psi(x)] is missing. As printed, Eq. (7) is not a valid GAN-style objective and does not describe the adversarial distillation loss that the section attributes to ADD.
  3. [Consistency Trajectory Model (CTM)] The CTM boundary condition is inconsistent with the stated parameterization. The text says the boundary condition is f_phi(x_t, t, 1) = x_1, but the parameterization f_phi(x_t, t, s) = (s/t) x_t + (1-s)/t F_phi(x_t, t, s) yields f_phi(x_t, t, t) = x_t and f_phi(x_t, t, 1) = (1/t)x_t + (1-1/t)F_phi, which never equals x_1. The actual CTM boundary condition is f_phi(x, t, t) = x. This error undermines the description of CTM and the subsequent discussion of SCoT, which builds on CTM's projection function.
  4. [Figure 1 and Fisher Divergence Loss section] The taxonomy is not exclusive as presented. SiDA appears in both the Fisher divergence / fidelity branch and the adversarial loss branch of Figure 1, and the text separately describes SiDA as a fidelity method and as an adversarial method. Since the survey claims to divide methods into three categories, a method appearing in two categories requires explicit explanation (e.g., as a combined method) or the taxonomy must be revised; otherwise the systematic categorization is ambiguous.
minor comments (4)
  1. [Variational Score Distillation (VSD)] The VSD gradient equation has a stray 'y' at the end of the displayed formula, and the equation uses x_0 where the one-step generator output f_phi(z) is meant; this makes the gradient expression confusing.
  2. [Straight Consistent Trajectory (SCoT)] The sentence 'The loss function of TraFlow considers the factors of soft consistency matching and trajectory straightness' introduces 'TraFlow' without definition; this appears to be a typo for SCoT, but as written it disrupts the description.
  3. [Preliminary, score function] The score function expression nabla_{x_t} log p_real(x_t) = -sigma_t^{-1}(x_t - alpha_t x_0) is dimensionally inconsistent; the correct conditional score is -sigma_t^{-2}(x_t - alpha_t x_0) (or - (x_t - alpha_t x_0)/sigma_t^2).
  4. [Various typographical errors] There are several typographical errors, including 'optmized' in the Diffusion-GAN paragraph, 'Score implit Matching' in the SiD paragraph, and 'qiang liu' in the reference list; these should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: survey organizes existing methods under a descriptive taxonomy and fits no parameters; equation-level inaccuracies are fidelity/correctness issues, not circular reductions.

full rationale

This paper is a literature survey, not a derivation or prediction pipeline. It fits no parameters, trains no model, and reports no experimental result of its own; consequently none of the rubric patterns applies: there is no self-definition, no fitted input renamed as prediction, no load-bearing self-citation, no imported uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known result presented as derivation. The central contribution is a three-way taxonomy (fidelity distillation, trajectory distillation, adversarial distillation), which is an organizational scheme and not a claim derived from its own premises. The only author self-citation, SCoT (Wu et al. 2025b), appears as one surveyed method and is described descriptively; it is not used to justify the taxonomy, to exclude alternative frameworks, or to ground any theorem, so it is not load-bearing. The substantive weaknesses are correctness/fidelity issues rather than circularity: Eq. (3) misstates the DDIM update by failing to scale the first bracket by sqrt(alpha_{t-1}/alpha_t); Eq. (7) omits the real-image expectation in the ADD GAN objective; the CTM paragraph states the boundary condition as f_phi(x_t,t,1)=x_1 while the given parameterization f_phi(x_t,t,s)=(s/t)x_t+(1-s/t)F_phi(x_t,t,s) yields f_phi(x_t,t,t)=x_t and never x_1 (the actual CTM boundary is f(x,t,t)=x); and "TraFlow" appears in the SCoT paragraph without definition. These flaws undercut the survey's promised faithful unified notation for comparing methods, but they are not reductions of an output to an input, so the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The survey introduces no free parameters and no invented entities. Its main unstated premises are the reliability of the teacher model, the standard diffusion and DDIM equations, and the validity of the taxonomic partition. The taxonomy assumption is the most fragile because the paper's own text violates it by overlapping categories.

assumptions (4)
  • domain assumption The pre-trained teacher model provides an accurate noise score estimator epsilon_theta(x_t,t) or velocity estimator v_theta(x_t,t).
    Every distillation method in the survey transfers knowledge from a teacher model; if the teacher is inaccurate, the survey's descriptions of distillation objectives lose their grounding. This is implicit in the setup and used throughout the fidelity and trajectory sections.
  • standard math The score function of the noise-corrupted data equals -sigma_t^{-1}(x_t - alpha_t x_0).
    This standard score expression is used in the Preliminary section to define the diffusion training loss (Eq. 2) and to justify the denoising formulas. It is a background result from DDPM and score-based generative modeling.
  • domain assumption The intended DDIM update is a valid deterministic sampler and the formula in Eq. (3), once corrected, is the one used by later methods.
    The survey relies on DDIM for SDI, consistency distillation, and trajectory comparisons, but the printed Eq. (3) is missing the parentheses around (x_t - sqrt(1-alpha_t)epsilon_theta)/sqrt(alpha_t). The assumed valid standard DDIM formula is needed for the later method descriptions.
  • ad hoc to paper The three-way taxonomy of fidelity, trajectory, and adversarial distillation is exhaustive and mutually exclusive.
    This is the survey's central organizing claim, introduced in the Overview and Figure 1. No argument is given that every method falls into exactly one bucket, and the text itself places SiDA in two buckets while describing adversarial loss as an orthogonal plug-in, so the partition assumption is fragile.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Survey on Pre-Trained Diffusion Model Distillations." pith.science (2026). https://pith.science/paper/DSCQJXP7

@misc{pith2026250208364,
  author       = {Pith},
  title        = {Pith review of: A Survey on Pre-Trained Diffusion Model Distillations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DSCQJXP7}},
  note         = {Machine review of arXiv:2502.08364}
}
read the original abstract

Diffusion Models~(DMs) have emerged as the dominant approach in Generative Artificial Intelligence (GenAI), owing to their remarkable performance in tasks such as text-to-image synthesis. However, practical DMs, such as stable diffusion, are typically trained on massive datasets and thus usually require large storage. At the same time, many steps may be required, i.e., recursively evaluating the trained neural network, to generate a high-quality image, which results in significant computational costs during sample generation. As a result, distillation methods on pre-trained DM have become widely adopted practices to develop smaller, more efficient models capable of rapid, few-step generation in low-resource environment. When these distillation methods are developed from different perspectives, there is an urgent need for a systematic survey, particularly from a methodological perspective. In this survey, we review distillation methods through three aspects: output loss distillation, trajectory distillation and adversarial distillation. We also discuss current challenges and outline future research directions in the conclusion.

Figures

Figures reproduced from arXiv: 2502.08364 by the authors.

Figure 1
Figure 1. Taxonomy of Pre-trained Diffusion model distillation methods. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualisations on trajectory distillation methods. Grey arrows represent the trajectories of pre-trained models, whereas [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 50 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    S.; Boffi, N

    Albergo, M. S.; Boffi, N. M.; and Vanden-Eijnden, E. 2023. Stochastic I nterpolants: A U nifying F ramework for F lows and D iffusions. arXiv preprint arXiv:2303.08797

  4. [4]

    A.; Zhai, S.; Hu, S.; Zheng, D.; Talbott, W.; and Gu, E

    Berthelot, D.; Autef, A.; Lin, J.; Yap, D. A.; Zhai, S.; Hu, S.; Zheng, D.; Talbott, W.; and Gu, E. 2023. Tract: D enoising D iffusion M odels with T ransitive C losure T ime- D istillation. arXiv preprint arXiv:2303.04248

  5. [5]

    Dieleman, S. 2024. The P aradox of D iffusion D istillation

  6. [6]

    Esser, P.; Kulal, S.; Blattmann, A.; Entezari, R.; M \"u ller, J.; Saini, H.; Levi, Y.; Lorenz, D.; Sauer, A.; Boesel, F.; et al. 2024. Scaling R ectified R low T ransformers for H igh- R esolution I mage S ynthesis. ICML

  7. [7]

    H.; and Pons, J

    Evans, Z.; Carr, C.; Taylor, J.; Hawley, S. H.; and Pons, J. 2024. Fast T iming- C onditioned L atent A udio D iffusion. ICML

  8. [8]

    Frans, K.; Hafner, D.; Levine, S.; and Abbeel, P. 2025. One S tep D iffusion via S hortcut M odels. ICLR

Show all 66 references
  1. [9]

    Z.; and He, K

    Geng, Z.; Deng, M.; Bai, X.; Kolter, J. Z.; and He, K. 2025 a . Mean flows for one-step generative modeling. arXiv preprint arXiv:2505.13447

  2. [10]

    Geng, Z.; Pokle, A.; Luo, W.; Lin, J.; and Kolter, J. Z. 2025 b . Consistency M odels M ade E asy. ICLR

  3. [11]

    Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. 2014. Generative A dversarial N etworks. NeurIPS

  4. [12]

    Gu, J.; Zhai, S.; Zhang, Y.; Liu, L.; and Susskind, J. M. 2023. Boot: D ata-free D istillation of D enoising D iffusion M odels with B ootstrapping. ICML 2023

  5. [13]

    Ho, J.; Jain, A.; and Abbeel, P. 2020 a . Denoising D iffusion P robabilistic M odels. NeurIPS

  6. [14]

    Ho, J.; Jain, A.; and Abbeel, P. 2020 b . Denoising D iffusion P robabilistic M odels. NeurIPS

  7. [15]

    Huang, Y.; Huang, J.; Liu, Y.; Yan, M.; Lv, J.; Liu, J.; Xiong, W.; Zhang, H.; Chen, S.; and Cao, L. 2024. Diffusion M odel-based I mage E diting: A S urvey. arXiv preprint arXiv:2402.17525

  8. [16]

    Karras, T.; Aittala, M.; Aila, T.; and Laine, S. 2022. Elucidating the D esign S pace of D iffusion-based G enerative M odels. NeurIPS

  9. [17]

    Karras, T.; Aittala, M.; Lehtinen, J.; Hellsten, J.; Aila, T.; and Laine, S. 2024. Analyzing and I mproving the T raining D ynamics of D iffusion M odels. CVPR

  10. [18]

    Kim, D.; Lai, C.-H.; Liao, W.-H.; Murata, N.; Takida, Y.; Uesaka, T.; He, Y.; Mitsufuji, Y.; and Ermon, S. 2024. Consistency T rajectory M odels: L earning P robability F low ODE T rajectory of D iffusion. ICLR

  11. [19]

    Lee, S.; Lin, Z.; and Fanti, G. 2024. Improving the training of rectified flows. Advances in neural information processing systems, 37: 63082--63109

  12. [20]

    Liang, Y.; Yang, X.; Lin, J.; Li, H.; Xu, X.; and Chen, Y. 2024. Lucid D reamer: T owards H igh- F idelity T ext-to-3 D G eneration via I nterval S core M atching. CVPR

  13. [21]

    T.; Ben-Hamu, H.; Nickel, M.; and Le, M

    Lipman, Y.; Chen, R. T.; Ben-Hamu, H.; Nickel, M.; and Le, M. 2023. Flow M atching for G enerative M odeling. ICLR

  14. [22]

    Liu, H.; Chen, Z.; Yuan, Y.; Mei, X.; Liu, X.; Mandic, D.; Wang, W.; and Plumbley, M. D. 2023 a . Audioldm: T ext-to- A udio G eneration with L atent D iffusion M odels. ICML

  15. [23]

    Liu, X.; Gong, C.; and qiang liu. 2023. Flow S traight and F ast: L earning to G enerate and T ransfer D ata with R ectified F low. ICLR

  16. [24]

    Liu, X.; Zhang, X.; Ma, J.; Peng, J.; et al. 2023 b . Instaflow: O ne S tep is E nough for H igh-quality D iffusion-based T ext-to-image G eneration. ICLR

  17. [25]

    Lu, C.; and Song, Y. 2025. Simplifying, S tabilizing and S caling C ontinuous-time C onsistency M odels. ICLR

  18. [26]

    Luhman, E.; and Luhman, T. 2021. Knowledge D istillation in I terative G enerative M odels for I mproved S ampling S peed. arXiv preprint arXiv:2101.02388

  19. [27]

    Lukoianov, A.; Borde, H. S. d. O.; Greenewald, K.; Guizilini, V. C.; Bagautdinov, T.; Sitzmann, V.; and Solomon, J. 2024. Score D istillation via R eparametrized DDIM . arXiv preprint arXiv:2405.15891

  20. [28]

    Luo, W. 2023. A C omprehensive S urvey on K nowledge D istillation of D iffusion M odels. arXiv preprint arXiv:2304.04262

  21. [29]

    Luo, W.; Hu, T.; Zhang, S.; Sun, J.; Li, Z.; and Zhang, Z. 2024 a . Diff- I nstruct: A U niversal A pproach for T ransferring K nowledge F rom P re-trained D iffusion M odels. NeurIPS

  22. [30]

    Z.; and Qi, G.-j

    Luo, W.; Huang, Z.; Geng, Z.; Kolter, J. Z.; and Qi, G.-j. 2024 b . One- S tep D iffusion D istillation through S core I mplicit M atching. arXiv preprint arXiv:2410.16794

  23. [31]

    Meng, C.; Rombach, R.; Gao, R.; Kingma, D.; Ermon, S.; Ho, J.; and Salimans, T. 2023. On D istillation of G uided D iffusion M odels. CVPR

  24. [32]

    H.; and Tran, A

    Nguyen, T. H.; and Tran, A. 2024. Swiftbrush: O ne- S tep T ext-to- I mage D iffusion M odel with V ariational S core D istillation. CVPR

  25. [33]

    T.; and Mildenhall, B

    Poole, B.; Jain, A.; Barron, J. T.; and Mildenhall, B. 2023. Dreamfusion: Text-to-3 D using 2 D D iffusion. ICLR

  26. [34]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution I mage S ynthesis with L atent D iffusion M odels. CVPR

  27. [35]

    Salimans, T.; and Ho, J. 2022. Progressive D istillation for F ast S ampling of D iffusion M odels. ICLR

  28. [36]

    Sauer, A.; Boesel, F.; Dockhorn, T.; Blattmann, A.; Esser, P.; and Rombach, R. 2024. Fast H igh- R esolution I mage S ynthesis with L atent A dversarial D iffusion D istillation. SIGGRAPH

  29. [37]

    Sauer, A.; Lorenz, D.; Blattmann, A.; and Rombach, R. 2025. Adversarial D iffusion D istillation. ECCV

  30. [38]

    Shuai, X.; Ding, H.; Ma, X.; Tu, R.; Jiang, Y.-G.; and Tao, D. 2024. A S urvey of M ultimodal- G uided I mage E diting with T ext-to- I mage D iffusion M odels. arXiv preprint arXiv:2406.14555

  31. [39]

    Song, J.; Meng, C.; and Ermon, S. 2021. Denoising D iffusion I mplicit M odels. ICLR

  32. [40]

    Song, Y.; and Dhariwal, P. 2024. Improved T echniques for T raining C onsistency M odels. ICLR

  33. [41]

    Song, Y.; Dhariwal, P.; Chen, M.; and Sutskever, I. 2023. Consistency M odels. ICML

  34. [42]

    P.; Kumar, A.; Ermon, S.; and Poole, B

    Song, Y.; Sohl-Dickstein, J.; Kingma, D. P.; Kumar, A.; Ermon, S.; and Poole, B. 2021. Score-based G enerative M odeling through S tochastic D ifferential E quations. ICLR

  35. [43]

    Wang, Z.; Lu, C.; Wang, Y.; Bao, F.; Li, C.; Su, H.; and Zhu, J. 2024. Prolificdreamer: H igh- F idelity and D iverse T ext-to- 3D G eneration with V ariational S core D istillation. NeurIPS

  36. [44]

    Wang, Z.; Zheng, H.; He, P.; Chen, W.; and Zhou, M. 2023. Diffusion- GAN : T raining GAN s with D iffusion. ICLR

  37. [45]

    L.; Juergens, D.; Bennett, N

    Watson, J. L.; Juergens, D.; Bennett, N. R.; Trippe, B. L.; Yim, J.; Eisenach, H. E.; Ahern, W.; Borst, A. J.; Ragotte, R. J.; Milles, L. F.; et al. 2023. De novo design of protein structure and function with RF diffusion. Nature

  38. [46]

    E.; Yang, K

    Wu, K. E.; Yang, K. K.; van den Berg, R.; Alamdari, S.; Zou, J. Y.; Lu, A. X.; and Amini, A. P. 2024 a . Protein S tructure G eneration via F olding D iffusion. Nature Communications

  39. [47]

    Wu, Y.; Chen, K.; Zhang, T.; Hui, Y.; Berg-Kirkpatrick, T.; and Dubnov, S. 2023. Large-scale C ontrastive L anguage- A udio P retraining with F eature F usion and K eyword-to- C aption A ugmentation. ICASSP

  40. [48]

    Wu, Z.; and Cao, L. 2023. C2VAE: Gaussian C opula-based VAE D iffering D isentangled from C oupled R epresentations with C ontrastive P osterior. arXiv preprint arXiv:2309.13303

  41. [49]

    Wu, Z.; Cao, L.; and Qi, L. 2024. e VAE : E volutionary V ariational A utoencoder. IEEE Transactions on Neural Networks and Learning Systems

  42. [50]

    Wu, Z.; Fan, X.; and Cao, L. 2025. Prog D iffusion: P rogressively S elf-encoding D iffusion M odels. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, 1633--1644

  43. [51]

    Wu, Z.; Fan, X.; Li, J.; Zhao, Z.; Chen, H.; and Cao, L. 2024 b . Paramrel: Learning P arameter S pace R epresentation via P rogressively E ncoding B ayesian F low N etworks. arXiv preprint arXiv:2405.15268

  44. [52]

    Wu, Z.; Fan, X.; Li, J.; Zhao, Z.; Chen, H.; and Cao, L. 2025 a . SepDiff: Self-Encoding Parameter Diffusion for Learning Latent Semantics. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, 3273--3284

  45. [53]

    Wu, Z.; Fan, X.; Wu, H.; and Cao, L. 2025 b . SCoT: Straight Consistent Trajectories for Pre-Trained Diffusion Model Distillation. arXiv preprint arXiv:2502.16972

  46. [54]

    Wu, Z.; Zhang, Q.; Zhou, J.; Chen, H.; and Liu, Y. 2025 c . WAVAE: A Weakly Augmented Variational Autoencoder for Time Series Anomaly Detection. Information Fusion, 103462

  47. [55]

    P.; Hou, T.; Wu, Y

    Xie, S.; Xiao, Z.; Kingma, D. P.; Hou, T.; Wu, Y. N.; Murphy, K. P.; Salimans, T.; Poole, B.; and Gao, R. 2024. EM D istillation for O ne- S tep D iffusion M odels. NeurIPS

  48. [56]

    Xu, Y.; Zhao, Y.; Xiao, Z.; and Hou, T. 2024. UFOGen : Y ou F orward O nce L arge S cale T ext-to- I mage G eneration via D iffusion GAN s. CVPR

  49. [57]

    Yan, R.; Chen, Y.; and Wang, X. 2025. Consistent flow distillation for text-to-3d generation. arXiv preprint arXiv:2501.05445

  50. [58]

    Yang, L.; Zhang, Z.; Song, Y.; Hong, S.; Xu, R.; Zhao, Y.; Shao, Y.; Zhang, W.; Cui, B.; and Yang, M.-H. 2022. Diffusion M odels: A C omprehensive S urvey of M ethods and A pplications. arXiv preprint arXiv:2209.00796

  51. [59]

    Yang, L.; Zhang, Z.; Zhang, Z.; Liu, X.; Xu, M.; Zhang, W.; Meng, C.; Ermon, S.; and Cui, B. 2024. Consistency F low M atching: D efining S traight F lows with V elocity C onsistency. arXiv preprint arXiv:2407.02398

  52. [60]

    L.; De Bortoli, V.; Mathieu, E.; Doucet, A.; Barzilay, R.; and Jaakkola, T

    Yim, J.; Trippe, B. L.; De Bortoli, V.; Mathieu, E.; Doucet, A.; Barzilay, R.; and Jaakkola, T. 2023. SE(3) D iffusion M odel with A pplication to P rotein B ackbone G eneration. arXiv preprint arXiv:2302.02277

  53. [61]

    Yin, T.; Gharbi, M.; Park, T.; Zhang, R.; Shechtman, E.; Durand, F.; and Freeman, W. T. 2024 a . Improved D istribution M atching D istillation for F ast I mage S ynthesis. NeurIPS

  54. [62]

    T.; and Park, T

    Yin, T.; Gharbi, M.; Zhang, R.; Shechtman, E.; Durand, F.; Freeman, W. T.; and Park, T. 2024 b . O ne- S tep D iffusion with D istribution M atching D istillation. CVPR

  55. [63]

    Zhou, M.; Wang, Z.; Zheng, H.; and Huang, H. 2025 a . Long and S hort G uidance in S core I dentity D istillation for O ne- S tep T ext-to- I mage G eneration. ICLR

  56. [64]

    Zhou, M.; Zheng, H.; Gu, Y.; Wang, Z.; and Huang, H. 2025 b . Adversarial S core identity D istillation: R apidly S urpassing the T eacher in O ne S tep. ICLR

  57. [65]

    Zhou, M.; Zheng, H.; Wang, Z.; Yin, M.; and Huang, H. 2024. Score I dentity D istillation: E xponentially F ast D istillation of P retrained D iffusion M odels for O ne- S tep G eneration. ICML

  58. [66]

    Zhu, Y.; Liu, X.; and Liu, Q. 2025. Slimflow: T raining S maller O ne- S tep D iffusion M odels with R ectified F low. ECCV

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.