Pith. sign in

REVIEW 2 major objections 4 minor 74 references

D2PO: Optimizing Diffusion Samplers via Dynamic Preference

T0 review · 2 major / 4 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Dynamic preferences and a score-based energy let few-step diffusion samplers keep texture fidelity instead of collapsing under teacher regression.

desk verdict Solid, usable preference-based fix for low-NFE sampler schedules; theory is looser than the empirics, but the gains hold up. read the letter →

arxiv 2607.06609 v1 pith:74VBHSV5 submitted 2026-07-07 cs.LG cs.AI

Jinkyu Kim , Jinyoung Choi , Bohyung Han This is my paper · ORCID
classification cs.LGcs.AI
keywords diffusionmodelssampleroptimizationtimestepschedulesDirectPreferencescore-basedenergyclassifier-freeguidancefew-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

Standard ways of speeding up diffusion models force a cheap student sampler to copy a high-quality teacher by regression. When the student has far fewer steps, it can only match the teacher's coarse structure and loses fine texture, so image quality actually falls even as the teacher improves. This paper replaces that fixed teacher with a preference objective: at every training step the student is paired against a denser version of its own current schedule, and the preferred (denser) sample is used as the winner. Preferences are measured by an energy built from the pretrained score network itself, so the comparison sees both global structure and high-frequency detail. The result is a self-refining loop that reduces discretization error without being locked to a fixed residual, and the learned schedules and guidance weights produce sharper, more preferred images under tight step budgets.

What carries the argument

Score-based energy: the squared difference of noise predictions of the pretrained score network at randomly sampled noise levels, used as the energy of an EBM surrogate so that DPO log-ratio terms become tractable energy differences; combined with a dynamic preference pair whose winner is the same policy run on a denser (e.g. 2N) schedule.

What would settle it

Train D2PO and a strong regression baseline (e.g. LD3) at 4–5 steps on the same frozen backbone and prompts; if blind human preference, HPSv2 and Aesthetic scores do not favor D2PO, or if FID collapses while those metrics rise, the claim that the dynamic preference + score energy better aligns with perceptual quality fails.

Watch

Extended reading notes

Core claim

Under low-NFE constraints, modeling the deterministic sampler as an energy-based policy and optimizing it with dynamic Direct Preference Optimization against a self-refined denser trajectory yields sampling policies whose perceptual quality (HPS, Aesthetic, human preference) exceeds that of regression-based schedulers, because the student is never forced to match a static teacher beyond its capacity.

Load-bearing premise

The paper assumes that the noise-prediction distance computed by the pretrained score network is a faithful multi-scale stand-in for human perceptual preference, and that the triangle-inequality bound relating dynamic loss to true discretization error still holds for the practical solvers and degradation used.

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

2 major / 4 minor

Summary. The paper introduces D2PO, a Direct Preference Optimization framework for learning low-dimensional diffusion sampler parameters (timestep schedules S and per-step CFG weights ω) while keeping the pretrained score network frozen. It models the deterministic sampler as an energy-based surrogate whose energy is the Monte-Carlo noise-prediction distance (Eqs. 17–19) induced by the pretrained network, then constructs dynamic preference pairs in which the winner is the same policy run on a denser (2N) schedule and the loser is a degraded version of the student output. The resulting logistic objective is claimed to produce higher perceptual quality (HPSv2, Aesthetic, human votes) than regression-based schedulers (LD3, GITS, DMN) under low-NFE budgets, while remaining competitive on FID, because the student is aligned to a self-refining trajectory that reduces discretization error rather than to a fixed high-NFE teacher.

Significance. If the empirical gains hold, D2PO supplies a lightweight, orthogonal alternative to both black-box search and student-teacher regression for few-step sampling; it freezes the generative backbone and optimizes only a handful of continuous parameters, making it complementary to weight-space distillation or RL fine-tuning. The experimental suite is unusually thorough for the sub-area: three ODE solvers, COCO T2I, ImageNet-256 latent, InstaFlow, SD 3.5-Medium, AFHQv2, resource-matched LD3†, CLIP scores, and a blind user study. The score-based energy and the dynamic denser-schedule target are genuine technical contributions. The theoretical sketch in §4.5 is only approximate, yet the practical recipe is immediately usable and the ablations (Tab. 4) isolate the two novel ingredients.

major comments (2)
  1. [§4.5 Theoretical analysis / Eqs. 21–23 vs. Eq. 19 & Alg. 1] The lower-bound argument L_dyn ≳ (1−2^{-k})ε_true (Eqs. 21–23) treats L_dyn as the direct metric ρ(π_ϕ,π_ϕ′). The implemented loss (Eq. 19 and Alg. 1), however, forms the loser by xl=G(sg[x_ϕ]) (G a low-pass filter) and optimizes a relative DPO logistic against an EMA/copy reference. This surrogate is not identical to ρ(π_ϕ,π_ϕ′); the degradation operator and the reference terms introduce additional degrees of freedom whose effect on continuous-time discretization error is unanalyzed. Consequently the claimed “systematic reduction of discretization error” does not follow rigorously from the triangle inequality. Either derive a corresponding bound for the actual objective or reframe §4.5 as informal motivation and rest the mechanism claim more explicitly on the ablations.
  2. [§4.4 Dynamic preference / Alg. 1] The degradation operator G that produces the losing sample is mentioned only by example (“e.g., a low-pass filter”) and never specified (filter type, cutoff frequency, whether it is applied in pixel or latent space, etc.). Because G directly shapes every preference pair and is listed among the free design choices, its precise definition and a short sensitivity study are required for reproducibility and for assessing whether the reported gains are robust to the choice of G.
minor comments (4)
  1. [Eq. 19 / Alg. 2] In the practical objective (Eq. 19) the expectation over t is written, yet Alg. 2 samples a single t. Clarify whether the Monte-Carlo estimate uses one or multiple noise levels per preference pair and report the value used in all tables.
  2. [§5.1 / App. A.1] Hyper-parameter ranges for β, λ, base learning rates and the precise form of the linear interpolation that produces the 2N schedule are given only in the appendix; a short summary in the main experimental setup would improve self-contained readability.
  3. [Fig. 1] Fig. 1 caption claims “severe artifacts” for larger Δ, but the visual difference between Δ=3 and Δ=4 is modest; either strengthen the visual example or soften the language.
  4. [§3.1 / §4.4] Typographical: “preal” should be p_real (Eq. 1); “sg[·]” is introduced without definition in the main text (only in Alg. 1).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: dynamic self-refinement is intentional design justified by standard numerical analysis, with all claims validated on external frozen metrics and ablations.

full rationale

The paper's core derivation chain (EBM surrogate for deterministic policies o score-based energy from the frozen pretrained network o DPO logistic on energy differences o dynamic preference via denser 2N refinement of the current student) does not reduce any claimed result to its own inputs by construction. The energy (Eqs. 16–18) is taken directly from the external pretrained score/noise network and is never fitted to the reported HPS/Aesthetic/FID numbers. The dynamic teacher (Sec. 4.4, Alg. 1) is deliberately self-referential, but the supporting argument in §4.5 is a standard triangle-inequality lower bound that assumes only a positive convergence order k of the underlying ODE solver (cited to a textbook [49]); it does not import a uniqueness theorem or ansatz from the authors' prior work, nor does it redefine the evaluation metrics. All quantitative claims are measured against frozen external proxies (HPSv2, LAION Aesthetic, FID, CLIP, blind human votes) and against independent baselines (LD3, GITS, DMN) on held-out prompts; the ablations in Tab. 4 further isolate components without circular reuse of the same data. The theory–practice gap noted by the skeptic (L_dyn vs. the actual degraded-xl DPO loss) is a correctness concern, not circularity. No self-definitional loop, fitted-input-as-prediction, load-bearing self-citation, or renamed known result is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim rests on the standard DPO closed-form, the modeling of a deterministic ODE sampler as an EBM whose energy is the pretrained score discrepancy, the assumption that a denser schedule of the same policy is a strictly better preference target, and a handful of hand-chosen optimization hyper-parameters. No new physical entities are postulated; the free parameters are ordinary ML knobs.

free parameters (4)
  • DPO temperature β = {10,50,100}
    Chosen from {10,50,100}; learning rates are scaled by 1/β. Directly multiplies the preference margin and therefore affects which schedules are selected.
  • EMA momentum λ for reference CFG weights
    Controls how quickly the reference policy tracks the student; not derived, set by hand.
  • learning-rate base values and gradient-accumulation steps
    Hand-tuned (1e-3/β for S1, 2e-4/(β N) for S2/ω, accumulation=4). Affect convergence of the low-dimensional policy.
  • degradation operator G for losing sample
    Low-pass filter (or equivalent) applied to the student output; its precise form is not ablated and influences the preference margin.
assumptions (4)
  • standard math The optimal policy of the KL-regularized reward maximization problem admits the closed-form log-ratio expression used by DPO (Eq. 8–9).
    Imported from Rafailov et al. 2023; used without re-derivation to obtain the D2PO logistic loss.
  • domain assumption A deterministic ODE sampler can be replaced by a smooth EBM surrogate whose energy is a distance to the sampler output, so that partition functions cancel in preference ratios (Eq. 12–14).
    Standard functional relaxation; correctness of gradients for the true Dirac policy is assumed rather than proved.
  • domain assumption Numerical ODE solvers of order k>0 reduce local truncation error by approximately 2^{-k} under 2× step refinement (Eq. 22).
    Invoked in §4.5 to claim that the dynamic teacher has strictly lower true error; classical numerical-analysis fact applied to the particular solvers used.
  • ad hoc to paper The noise-prediction distance integrated over t (with weight σ_t^{2}) is a faithful multi-scale surrogate for perceptual discrepancy.
    Core modeling choice of §4.3; justified by score geometry but not independently validated against human ratings outside the final experiments.
invented entities (2)
  • score-based energy for the EBM sampler surrogate
    purpose: Provides a differentiable, multi-scale preference signal that does not require an external reward model or LPIPS network.
    Defined in Eqs. 16–18; independent evidence is only the ablation showing it outperforms LPIPS, not an external falsifiable prediction.
  • dynamic denser-schedule preference target (ϕ′)
    purpose: Replaces a fixed external teacher so that the preference signal improves with the student and avoids a residual error floor.
    Constructed by linear interpolation of the current schedule to 2N steps; the self-improving claim rests on this construction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of D2PO: Optimizing Diffusion Samplers via Dynamic Preference." pith.science (2026). https://pith.science/paper/74VBHSV5

@misc{pith2026260706609,
  author       = {Pith},
  title        = {Pith review of: D2PO: Optimizing Diffusion Samplers via Dynamic Preference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/74VBHSV5}},
  note         = {Machine review of arXiv:2607.06609}
}
read the original abstract

We propose D2PO (Dynamic Direct Preference Optimization), a principled framework for optimizing diffusion sampling policies with respect to timestep schedules and classifier-free guidance (CFG) weights. Our work is motivated by a fundamental limitation of existing student-teacher regression frameworks; low-NFE student samplers are trained to mimic high-NFEteachers, often sacrificing high-frequency texture fidelity while preserving coarse global structures, thereby misaligning the sampler with perceptual quality. D2PO addresses this challenge by reformulating sampler optimization as a preference-based alignment problem, leveraging the Direct Preference Optimization (DPO) framework. To make DPO applicable to diffusion samplers, we model the sampling policy as an energy-based model (EBM), transforming preference comparisons into tractable energy differences. We further introduce a novel energy formulation derived directly from the pretrained score network, enabling preference evaluation in perturbed spaces that jointly capture structural consistency and fine-grained details. Moreover, we introduce dynamic preferences, where the preferred samples used for alignment progressively improve as the sampling policies are learned. This self-improving mechanism replaces rigid static teacher supervision with an iterative, preference-guided refinement process, providing progressively stronger alignment signals. Extensive experiments demonstrate that D2PO aligns diffusion samplers with perceptual quality more faithfully, unlocking the full potential of high-quality teachers and consistently outperforming conventional regression-based schedulers under low-NFE constraints.

Figures

Figures reproduced from arXiv: 2607.06609 by the authors.

Figure 1
Figure 1. Qualitative evidence of the performance bottleneck in LD3 [50]. All images are generated by the same model with NFE = 4. The columns show the impact of increasing the NFE gap (∆ = T − S) between the teacher (T) and the student (S). While a small gap (∆ = 1) yields high-quality outputs, larger gaps (up to ∆ = 4) lead to severe artifacts, demonstrating LD3’s inability to leverage high-fidelity teachers. computational … view at source ↗
Figure 2
Figure 2. Qualitative comparison of various discretization methods on Stable Diffusion v1.5 using iPNDM across different number of time steps (4 to 7). D2PO consistently produces sharper details and fewer artifacts compared to baselines. 5.4 Qualitative results [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 74 canonical work pages

  1. [1]

    In: ICLR (2022)

    Bao, F., Li, C., Zhu, J., Zhang, B.: Analytic-DPM: An analytic estimate of the optimal reverse variance in diffusion probabilistic models. In: ICLR (2022)

    work page 2022

  2. [2]

    In: ICLR (2018)

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD GANs. In: ICLR (2018)

    work page 2018

  3. [3]

    In: ICLR (2024)

    Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. In: ICLR (2024)

    work page 2024

  4. [4]

    In: ICLR (2019)

    Brock, A., Donahue, J., Simonyan, K.: Large scale GAN training for high fidelity natural image synthesis. In: ICLR (2019)

    work page 2019

  5. [5]

    In: ICML (2024)

    Chen, D., Zhou, Z., Wang, C., Shen, C., Lyu, S.: On the trajectory regularity of ODE-based diffusion sampling. In: ICML (2024)

    work page 2024

  6. [6]

    Training Deep Nets with Sublinear Memory Cost

    Chen, T., Xu, B., Zhang, C., Guestrin, C.: Training deep nets with sublinear mem- ory cost. arXiv preprint arXiv:1604.06174 (2016)

    work page Pith review arXiv 2016

  7. [7]

    In: The 40th Conference on Uncertainty in Artificial Intelligence (2024)

    Chen, Y., He, F., Fu, S., Tian, X., Tao, D.: Adaptive time-stepping schedules for diffusion models. In: The 40th Conference on Uncertainty in Artificial Intelligence (2024)

    work page 2024

  8. [8]

    In: ICML

    Chen, Z., Deng, Y., Yuan, H., Ji, K., Gu, Q.: Self-play fine-tuning converts weak language models to strong language models. In: ICML. PMLR (2024)

    work page 2024

Show all 74 references
  1. [9]

    In: ICLR (2025)

    Choi, J., Kang, J., Han, B.: Enhanced diffusion sampling via extrapolation with multiple ode solutions. In: ICLR (2025)

    2025

  2. [10]

    In: NeurIPS (2021)

    Dhariwal, P., Nichol, A.: Diffusion models beat GANs on image synthesis. In: NeurIPS (2021)

    2021

  3. [11]

    In: NeurIPS (2022)

    Dockhorn, T., Vahdat, A., Kreis, K.: GENIE: Higher-order denoising diffusion solvers. In: NeurIPS (2022)

    2022

  4. [12]

    arXiv preprint arXiv:2403.03206 (2024)

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., et al.: Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206 (2024)

    2024 arXiv

  5. [13]

    In: NeurIPS (2023)

    Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Reinforcement learning for fine-tuning text- to-image diffusion models. In: NeurIPS (2023)

    2023

  6. [14]

    In: ICML (2025)

    Frankel, E., Chen, S., Li, J., Koh, P.W., Ratliff, L.J., Oh, S.: S4S: Solving for a fast diffusion model solver. In: ICML (2025)

    2025

  7. [15]

    arXiv preprint arXiv:2510.00815 (2025)

    Galashov, A., Pokle, A., Doucet, A., Gretton, A., Delbracio, M., De Bortoli, V.: Learn to guide your diffusion model. arXiv preprint arXiv:2510.00815 (2025)

    2025

  8. [16]

    In: ICML (2018)

    Haarnoja, T., Zhou, A., Abbeel, P., Levine, S.: Soft actor-critic: Off-policy maxi- mum entropy deep reinforcement learning with a stochastic actor. In: ICML (2018)

    2018

  9. [17]

    arXiv preprint arXiv:1604.00772 (2016)

    Hansen, N.: The CMA evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772 (2016)

    2016 arXiv

  10. [18]

    In: NeurIPS (2017)

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)

    2017

  11. [19]

    In: NeurIPS (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)

    2020

  12. [20]

    arXiv preprint arXiv:2207.12598 (2022)

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

    2022 arXiv

  13. [21]

    In: NeurIPS (2022)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: NeurIPS (2022)

    2022

  14. [22]

    In: ICLR (2017) D2PO: Optimizing Diffusion Samplers via Dynamic Preference 17

    Jang, E., Gu, S., Poole, B.: Categorical reparameterization with Gumbel-Softmax. In: ICLR (2017) D2PO: Optimizing Diffusion Samplers via Dynamic Preference 17

    2017

  15. [23]

    In: CVPR (2024)

    Kang, J., Choi, J., Choi, S., Han, B.: Observation-guided diffusion probabilistic models. In: CVPR (2024)

    2024

  16. [24]

    In: NeurIPS (2022)

    Karras,T.,Aittala,M.,Aila,T.,Laine,S.:Elucidatingthedesignspaceofdiffusion- based generative models. In: NeurIPS (2022)

    2022

  17. [25]

    In: ICLR (2024)

    Kim, D., Lai, C.H., Liao, W.H., Murata, N., Takida, Y., Uesaka, T., He, Y., Mitsu- fuji, Y., Ermon, S.: Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In: ICLR (2024)

    2024

  18. [26]

    In: NeurIPS (2021)

    Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. In: NeurIPS (2021)

    2021

  19. [27]

    NeurIPS (2018)

    Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolu- tions. NeurIPS (2018)

    2018

  20. [28]

    In: ICCV (2023)

    Li, L., Li, H., Zheng, X., Wu, J., Xiao, X., Wang, R., Zheng, M., Pan, X., Chao, F., Ji, R.: AutoDiffusion: Training-free optimization of time steps and architectures for automated diffusion model acceleration. In: ICCV (2023)

    2023

  21. [29]

    In: CVPR (2025)

    Liang, Z., Yuan, Y., Gu, S., Chen, B., Hang, T., Cheng, M., Li, J., Zheng, L.: Aesthetic post-training diffusion models from generic preferences with step-by-step preference optimization. In: CVPR (2025)

    2025

  22. [30]

    In: ECCV (2014)

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: ECCV (2014)

    2014

  23. [31]

    In: NeurIPS (2025)

    Liu, J., Liu, G., Liang, J., Li, Y., Liu, J., Wang, X., Wan, P., Zhang, D., Ouyang, W.: Flow-GRPO: Training flow matching models via online rl. In: NeurIPS (2025)

    2025

  24. [32]

    In: ICLR (2022)

    Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: ICLR (2022)

    2022

  25. [33]

    In: ICLR (2024)

    Liu, X., Zhang, X., Ma, J., Peng, J., Liu, Q.: InstaFlow: One step is enough for high-quality diffusion-based text-to-image generation. In: ICLR (2024)

    2024

  26. [34]

    In: NeurIPS (2022)

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: NeurIPS (2022)

    2022

  27. [35]

    arXiv preprint arXiv:2211.01095 (2022)

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022)

    2022 arXiv

  28. [36]

    NeurIPS (2024)

    Ma, X., Fang, G., Bi Mi, M., Wang, X.: Learning-to-cache: Accelerating diffusion transformer via layer caching. NeurIPS (2024)

    2024

  29. [37]

    NeurIPS (2022)

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. NeurIPS (2022)

    2022

  30. [38]

    In: NeurIPS (2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. In: NeurIPS (2019)

    2019

  31. [39]

    In: NeurIPS (2023)

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Di- rect preference optimization: Your language model is secretly a reward model. In: NeurIPS (2023)

    2023

  32. [40]

    In: CVPR (2022)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)

    2022

  33. [41]

    ICML (2024)

    Sabour, A., Fidler, S., Kreis, K.: Align your steps: Optimizing sampling schedules in diffusion models. ICML (2024)

    2024

  34. [42]

    In: ICLR (2022)

    Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2022)

    2022

  35. [43]

    In: NeurIPS (2024)

    Salimans, T., Mensink, T., Heek, J., Hoogeboom, E.: Multistep distillation of dif- fusion models via moment matching. In: NeurIPS (2024)

    2024

  36. [44]

    Kim et al

    Schuhmann, C.: Laion-aesthetics.https://laion.ai/blog/laion- aesthetics/ (2022), accessed: 2023-11-10 18 J. Kim et al

    2022

  37. [45]

    In: ICLR (2023)

    Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-A-Video: Text- to-video generation without text-video data. In: ICLR (2023)

    2023

  38. [46]

    In: ICML (2015)

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: ICML (2015)

    2015

  39. [47]

    In: ICML (2023)

    Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: ICML (2023)

    2023

  40. [48]

    In: ICLR (2021)

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: ICLR (2021)

    2021

  41. [49]

    Cambridge University Press, 1 edn

    Süli, E., Mayers, D.: An Introduction to Numerical Analysis. Cambridge University Press, 1 edn. (2003)

    2003

  42. [50]

    In: ICLR (2025)

    Tong, V., Trung-Dung, H., Liu, A., Broeck, G.V.d., Niepert, M.: Learning to dis- cretize denoising diffusion ODEs. In: ICLR (2025)

    2025

  43. [51]

    In: NeurIPS (2021)

    Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. In: NeurIPS (2021)

    2021

  44. [52]

    In: CVPR (2024)

    Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Purushwalkam, S., Ermon, S., Xiong, C., Joty, S., Naik, N.: Diffusion model alignment using direct preference optimization. In: CVPR (2024)

    2024

  45. [53]

    In: CVPR (2026)

    Wang, F.Y., Zhou, H., Yuan, L., Woo, S., Gong, B., Han, B., Yang, M.H., Zhang, H., Zhu, Y., Liu, T., Zhao, L.: Image diffusion preview with consistency solver. In: CVPR (2026)

    2026

  46. [54]

    arXiv preprint arXiv:2308.06571 (2023)

    Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: ModelScope text- to-video technical report. arXiv preprint arXiv:2308.06571 (2023)

    2023 arXiv

  47. [55]

    arXiv preprint arXiv:2305.16213 (2023)

    Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: ProlificDreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)

    2023 arXiv

  48. [56]

    In: ICLR (2022)

    Watson, D., Chan, W., Ho, J., Norouzi, M.: Learning fast samplers for diffusion models by differentiating through sample quality. In: ICLR (2022)

    2022

  49. [57]

    arXiv preprint arXiv:2106.03802 (2021)

    Watson, D., Ho, J., Norouzi, M., Chan, W.: Learning to efficiently sample from diffusion probabilistic models. arXiv preprint arXiv:2106.03802 (2021)

    2021 arXiv

  50. [58]

    arXiv preprint arXiv:2306.09341 (2023)

    Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 (2023)

    2023 arXiv

  51. [59]

    In: ICLR (2022)

    Xiao, Z., Kreis, K., Vahdat, A.: Tackling the generative learning trilemma with denoising diffusion GANs. In: ICLR (2022)

    2022

  52. [60]

    In: CVPR (2024)

    Xue, S., Liu, Z., Chen, F., Zhang, S., Hu, T., Xie, E., Li, Z.: Accelerating diffusion sampling with optimized time steps. In: CVPR (2024)

    2024

  53. [61]

    In: CVPR (2024)

    Yang, K., Tao, J., Lyu, J., Ge, C., Chen, J., Shen, W., Zhu, X., Li, X.: Using human feedback to fine-tune diffusion models without any reward model. In: CVPR (2024)

    2024

  54. [62]

    NeurIPS (2024)

    Ye, H., Yuan, J., Xia, R., Yan, X., Chen, T., Yan, J., Shi, B., Zhang, B.: Training- free adaptive diffusion with bounded difference approximation strategy. NeurIPS (2024)

    2024

  55. [63]

    In: NeurIPS (2024)

    Yin, T., Gharbi, M., Park, T., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T.: Improved distribution matching distillation for fast image synthesis. In: NeurIPS (2024)

    2024

  56. [64]

    In: CVPR (2024)

    Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: CVPR (2024)

    2024

  57. [65]

    NeurIPS (2024) D2PO: Optimizing Diffusion Samplers via Dynamic Preference 19

    Yuan, H., Chen, Z., Ji, K., Gu, Q.: Self-play fine-tuning of diffusion models for text-to-image generation. NeurIPS (2024) D2PO: Optimizing Diffusion Samplers via Dynamic Preference 19

    2024

  58. [66]

    In: CVPR (2023)

    Zhang, G., Kenta, N., Kleijn, W.B.: Lookahead diffusion probabilistic models for refining mean estimation. In: CVPR (2023)

    2023

  59. [67]

    In: ICLR (2024)

    Zhang, G., Kenta, N., Kleijn, W.B.: On accelerating diffusion-based sampling pro- cess via improved integration approximation. In: ICLR (2024)

    2024

  60. [68]

    In: ICLR (2023)

    Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: ICLR (2023)

    2023

  61. [69]

    In: CVPR (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)

    2018

  62. [70]

    In: NeurIPS (2023)

    Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: UniPC: A unified predictor-corrector framework for fast sampling of diffusion models. In: NeurIPS (2023)

    2023

  63. [71]

    arXiv preprint arXiv:2402.19159 (2024)

    Zheng, J., Hu, M., Fan, Z., Wang, C., Ding, C., Tao, D., Cham, T.J.: Trajectory consistency distillation. arXiv preprint arXiv:2402.19159 (2024)

    2024 arXiv

  64. [72]

    arXiv preprint arXiv:2211.11018 (2022)

    Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018 (2022)

    2022 arXiv

  65. [73]

    In: ICLR (2025)

    Zhou, M., Zheng, H., Gu, Y., Wang, Z., Huang, H.: Adversarial score identity distillation: Rapidly surpassing the teacher in one step. In: ICLR (2025)

    2025

  66. [74]

    An old photo of an umbrella and chairs at the beach

    Zhou, M., Zheng, H., Wang, Z., Yin, M., Huang, H.: Score identity distillation: Ex- ponentially fast distillation of pretrained diffusion models for one-step generation. In: ICML (2024) 20 J. Kim et al. A Experimental Setup A.1 Implementation details We implement our method us...

    2024

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.