Pith. sign in

REVIEW 2 major objections 4 minor 74 references

Dynamic preferences and a score-based energy let few-step diffusion samplers keep texture fidelity instead of collapsing under teacher regression.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 01:55 UTC pith:74VBHSV5

load-bearing objection Solid, usable preference-based fix for low-NFE sampler schedules; theory is looser than the empirics, but the gains hold up. the 2 major comments →

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

D2PO: Optimizing Diffusion Samplers via Dynamic Preference

Jinkyu Kim , Jinyoung Choi , Bohyung Han This is my paper
classification cs.LG cs.AI
keywords diffusion modelssampler optimizationtimestep schedulesDirect Preference Optimizationscore-based energyclassifier-free guidancefew-step generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

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.

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.

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.

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.

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 this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, 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

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.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 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.
axioms (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 no independent evidence
    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′) no independent evidence
    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.

pith-pipeline@v1.1.0-grok45 · 24340 in / 3209 out tokens · 43216 ms · 2026-07-11T01:55:30.682404+00:00 · methodology

0 comments
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 Bohyung Han, Jinkyu Kim, Jinyoung Choi.

Figure 1
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. comp… view at source ↗
Figure 2
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)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

74 extracted references · 74 canonical work pages · 11 internal anchors

  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 internal anchor Pith review Pith/arXiv 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

  9. [9]

    In: ICLR (2025)

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

    work page 2025

  10. [10]

    In: NeurIPS (2021)

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

    work page 2021

  11. [11]

    In: NeurIPS (2022)

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

    work page 2022

  12. [12]

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [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)

    work page 2023

  14. [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)

    work page 2025

  15. [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)

    work page arXiv 2025

  16. [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)

    work page 2018

  17. [17]

    The CMA Evolution Strategy: A Tutorial

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

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  18. [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)

    work page 2017

  19. [19]

    In: NeurIPS (2020)

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

    work page 2020

  20. [20]

    Classifier-Free Diffusion Guidance

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

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  21. [21]

    In: NeurIPS (2022)

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

    work page 2022

  22. [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

    work page 2017

  23. [23]

    In: CVPR (2024)

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

    work page 2024

  24. [24]

    In: NeurIPS (2022)

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

    work page 2022

  25. [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)

    work page 2024

  26. [26]

    In: NeurIPS (2021)

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

    work page 2021

  27. [27]

    NeurIPS (2018)

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

    work page 2018

  28. [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)

    work page 2023

  29. [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)

    work page 2025

  30. [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)

    work page 2014

  31. [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)

    work page 2025

  32. [32]

    In: ICLR (2022)

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

    work page 2022

  33. [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)

    work page 2024

  34. [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)

    work page 2022

  35. [35]

    DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  36. [36]

    NeurIPS (2024)

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

    work page 2024

  37. [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)

    work page 2022

  38. [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)

    work page 2019

  39. [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)

    work page 2023

  40. [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)

    work page 2022

  41. [41]

    ICML (2024)

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

    work page 2024

  42. [42]

    In: ICLR (2022)

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

    work page 2022

  43. [43]

    In: NeurIPS (2024)

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

    work page 2024

  44. [44]

    Kim et al

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

    work page 2022

  45. [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)

    work page 2023

  46. [46]

    In: ICML (2015)

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

    work page 2015

  47. [47]

    In: ICML (2023)

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

    work page 2023

  48. [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)

    work page 2021

  49. [49]

    Cambridge University Press, 1 edn

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

    work page 2003

  50. [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)

    work page 2025

  51. [51]

    In: NeurIPS (2021)

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

    work page 2021

  52. [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)

    work page 2024

  53. [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)

    work page 2026

  54. [54]

    ModelScope Text-to-Video Technical Report

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

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  55. [55]

    ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillation

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  56. [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)

    work page 2022

  57. [57]

    Learning to Efficiently Sample from Diffusion Probabilistic Models

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

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  58. [58]

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

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  59. [59]

    In: ICLR (2022)

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

    work page 2022

  60. [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)

    work page 2024

  61. [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)

    work page 2024

  62. [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)

    work page 2024

  63. [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)

    work page 2024

  64. [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)

    work page 2024

  65. [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

    work page 2024

  66. [66]

    In: CVPR (2023)

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

    work page 2023

  67. [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)

    work page 2024

  68. [68]

    In: ICLR (2023)

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

    work page 2023

  69. [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)

    work page 2018

  70. [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)

    work page 2023

  71. [71]

    Trajectory Consistency Distillation: Improved Latent Consistency Distillation by Semi-Linear Consistency Function with Trajectory Mapping

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  72. [72]

    MagicVideo: Efficient Video Generation With Latent Diffusion Models

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  73. [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)

    work page 2025

  74. [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 using PyTorch [38] and adopt the pretrained, frozen Stable Diffusion v1.5 [40] as ...

    work page arXiv 2024