Pith. sign in

REVIEW 5 major objections 6 minor 24 references

Differentiable Solver Search for Fast Diffusion Sampling

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A differentiable search over solver coefficients and timesteps yields a fast diffusion sampler that beats hand-designed multi-step solvers, with FID 2.33 on DiT-XL/2 at 10 steps.

desk verdict A genuinely useful reframing of diffusion solver design with strong FID tables, but the searched-loss-to-FID link is load-bearing and the paper itself says the two are misaligned. read the letter →

arxiv 2505.21114 v1 pith:DQ35AC5X submitted 2025-05-27 cs.CV

Shuai Wang , Zexian Li , Qipeng zhang , Tianhui Song , Xubin Li , Tiezheng Ge , Bo Zheng , Limin Wang This is my paper
classification cs.CV
keywords diffusionmodelsfastsamplingODEsolverssolversearchdifferentiableoptimizationtimesteprectifiedflowDDPM
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to establish that the interpolation form used in Adams-like diffusion solvers is not what matters; after pre-integration, any multistep update collapses into a weighted sum of previous velocity predictions. It argues that the only quantities worth optimizing are the solver coefficients and the sampling timesteps, and it introduces a differentiable search procedure to find them. On ImageNet 256 at 10 steps, the searched solver reaches FID 2.33 on DiT-XL/2, 2.40 on SiT-XL/2, and 2.35 on FlowDCN-XL/2, beating DPM-Solver++ and UniPC at matched function evaluations. The paper further claims the searched solver transfers across architectures, resolutions, and model sizes without retraining the generative model.

What carries the argument

The load-bearing object is the compact search space of solver coefficients $c_j^i$ and timesteps $t_i$ in the update $x_{i+1} \approx x_i + \sum_{j=0}^{i} v_j c_j^i (t_{i+1}-t_i)$. The argument shows that any universal interpolation function, once integrated over $[t_i, t_{i+1}]$, collapses into constants, and that the $x$-dependent coefficient should be replaced by its expectation over $x_i$. The search parametrizes timestep deltas with a softmax over unbounded logits and the coefficient matrix as lower-triangular with its diagonal constrained so that $\sum_k c_k^i = 1$, then optimizes a mono-alignment MSE loss to an $L$-step Euler reference trajectory plus a Huber auxiliary loss.

What would settle it

Reproduce Tables 2 and 4 under identical checkpoints, CFG settings, and evaluation code, varying only the reference-trajectory length $L$ in the mono-alignment loss: if the 10-step FID of 2.33 does not fall below DPM-Solver++ and UniPC at matched NFE, or if the gains disappear when $L$ is changed, the paper's central claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that $t$-related Lagrange interpolation is suboptimal for diffusion models, and that a universal interpolation function reduces, through pre-integration, to a set of coefficients $c_j^i$. Replacing the $x$-dependent pre-integral coefficient $C_j(x_i)$ with its expectation $\mathbb{E}_{x_i}[C_j(x_i)]$ yields scheduler-specific constants that preserve generalization. Optimizing these coefficients together with the timesteps $t_i$ produces a multistep solver that, at 10 steps, achieves FID 2.33 on DiT-XL/2, 2.40 on SiT-XL/2, and 2.35 on FlowDCN-XL/2 on ImageNet 256, outperforming DPM-Solver++ and UniPC at the same NFE. The paper also claims that with an optimized high-order solver, DDPM/VP models can match or surpass rectified-flow models under similar step budgets.

Load-bearing premise

The method hinges on treating the mean squared error to a long Euler sampling path as a faithful stand-in for final image quality, even though the paper's own ablations show that reconstruction error and FID can disagree.

Editorial extensions

If this is right

  • At 10 steps on ImageNet 256, the searched solver gives FID 2.33 on DiT-XL/2, 2.40 on SiT-XL/2, and 2.35 on FlowDCN-XL/2, all below the reported DPM-Solver++ and UniPC values at the same NFE.
  • A solver searched once on a small model transfers to larger models and to 512x512 resolution, so the search does not need to be rerun for each deployment.
  • The same searched solver applies to text-to-image rectified-flow models such as Flux.1-dev and SD3, and to DDPM-family models sharing the DiT noise schedule, improving visual quality from 5 to 10 steps.
  • Combining the searched solver with distillation yields 4-step FID 18.99 on SD1.5 and 11.4 on SDXL in the paper's COCO evaluations, indicating the search is compatible with existing acceleration methods.
  • The searched timestep patterns differ between rectified-flow and DDPM/VP, suggesting that optimal sampling schedules are model-family-specific rather than universal.

Reading between the lines

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

  • Because the paper's own ablations show a mismatch between FID and reconstruction error, a search that directly optimizes a perceptual or distribution-matching score on a small validation set could push 5-6 step FID further than mono-alignment does.
  • The transfer of a solver searched on FlowDCN-B/2 to SiT-XL/2 and to 512px models suggests that solver coefficients are more tied to the noise scheduler than to model architecture, so one search per noise schedule might serve an entire model family.
  • A natural testable extension is to search within a predictor-corrector structure; the paper's limitation discussion itself predicts this would improve large-CFG text-to-image generation.
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

5 major / 6 minor

Summary. The paper proposes a data-driven, differentiable search over solver coefficients and timestep schedules for fast diffusion sampling. It argues that the specific interpolation form in Adams-like multistep solvers is unimportant and that the effective degrees of freedom reduce to a compact space of coefficients M and timestep deltas. The method optimizes these parameters by aligning the sampled trajectory to an L-step Euler reference trajectory. Experiments report FID 2.33 on DiT-XL/2, 2.40 on SiT-XL/2, and 2.35 on FlowDCN-XL/2 at 10 steps on ImageNet 256, together with transfer results at 512 resolution and on PixArt/FLUX/SD3. The appendix contains searched parameter tables, ablations, and theoretical error-bound proofs.

Significance. If the empirical results hold, this is a practically valuable contribution: a no-retraining solver that consistently beats DPM-Solver++ and UniPC at matched NFE, and even outperforms some distillation methods at low step counts, using only 50,000 search samples. The paper's strength is the breadth of the evaluation across DiT, SiT, FlowDCN, PixArt, and multiple resolutions, and the release of full searched coefficient/timestep tables in Appendix J. The main caveat is that the search objective is a per-sample trajectory-alignment loss, whereas the reported metric is distribution-level FID; the paper itself acknowledges this mismatch in Appendix G. The theoretical optimality arguments also rely on strong assumptions that are not verified. The result is therefore plausible but conditional on additional validation.

major comments (5)
  1. [Appendix G and §5, Algorithm 2] The central empirical claim rests on the assumption that minimizing per-sample MSE/Huber alignment to an L-step Euler reference trajectory is a valid surrogate for final FID. Appendix G explicitly states that ablation studies reveal a mismatch between FID performance and reconstruction error, and Figure 3b shows that the searched solver's FID depends on the chosen reference step count L. The proposed remedy, incorporating distribution-matching supervision, is promised but not run. This is load-bearing: the reported gains over DPM-Solver++ and UniPC are only as strong as this unvalidated surrogate-to-quality link. Please provide the distribution-matching ablation, or otherwise empirically validate the surrogate, before the main claim can be accepted.
  2. [Appendix H, proof of Theorem 4.2] The proof of optimality in the pre-integral argument assumes "different vj are orthogonal and ||vj||^2 = 1" in order to solve the stationarity condition for the scalar coefficients a_j. No evidence is given that diffusion velocity fields along a sampling trajectory satisfy orthogonality or unit norm, and the assumption is not stated in Theorem 4.2 itself. Furthermore, the error expectation derived in Equations (22)-(27) is an L1 norm of a sum, while the optimization in Equations (28)-(29) minimizes a squared L2 objective; minimizing one does not imply minimality of the other without additional structure. The theorem should be restated as conditional on these assumptions, or the assumptions should be verified empirically on the models in Section 7.
  3. [Equations (34)-(43) and Theorem 4.5] The total error bound in Theorem 4.5 omits the interpolation remainder O(dx^m)+O(dt^n) with the statement that it is "much smaller" than the other term, but no quantitative estimate is provided. In the regime of interest, 5-10 steps on [0,1], dt is not small, and the dx term depends on sampled trajectories whose magnitude is not controlled by the stated assumptions. Without this estimate, the comparison in Theorem 4.5 does not establish that the searched solver has a smaller upper error bound than Adams-like multistep methods. This is a load-bearing part of the paper's theoretical justification for the search space.
  4. [Tables 2-5, §7.1-7.2] The reported FID values are single evaluations without error bars, multiple seeds, or released code. Several headline numbers are large margins, but some comparisons are small (e.g., DiT-XL/2-R512 at 10 steps, 3.64 vs. 3.81; PixArt at 10 steps, 33.4 vs. 35.6 in Table 5), and the paper gives no indication of evaluation variance. Since the central claim is a comparative improvement at matched NFE, please report variance across at least three evaluation runs or confirm that the standard ADM/FID protocol is followed with fixed seeds.
  5. [Appendix D, §7.2] Appendix D states that applying the searched solver to noise schedulers other than the one used during search yields meaningless results, and that DiT and SD1.5 are not compatible despite both being DDPM-style. This is an important boundary condition on the claimed generality. The paper currently emphasizes generality across architectures, resolutions, and model sizes; the scheduler-specific nature of the search should be stated more prominently in the abstract or introduction so that readers do not over-generalize the claim.
minor comments (6)
  1. [Abstract and §1] The abstract says "significantly outperforms traditional solvers (even some distillation methods)"; consider making clear that the comparison to distillation is limited to the specific FlowTurbo numbers in Table 1, not a general claim.
  2. [Assumption 4.1] The assumption text says "O(dtm) is the nth-order infinitesimal for dt"; this should be O(dt^n), and the notation in the displayed equation should match.
  3. [Appendix H, Equations (23) and (26)] There are unbalanced parentheses and missing norm delimiters in these equations, e.g., the expression "C_j(x_i) - b_j^i||" is missing an opening norm delimiter. Please proofread the appendix equations.
  4. [Appendix I, first paragraph] "We donate the continuous integral result" should be "We denote"; also "triangular inequalities" should be "triangle inequalities."
  5. [Figure 3] The subfigure captions "FID of Search Model" and "FID of RefTraj Steps" are unclear; define what "Search Model" and "RefTraj" mean and specify the reference trajectory step count L and the search model used in each panel.
  6. [Section 4.2, Equation (14)] The phrase "We simply use the first order Taylor expansion of x(t) at xi" is accurate, but the displayed equation omits the remainder term; adding the remainder explicitly would make the subsequent Assumption 4.1 easier to follow.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: searched coefficients and timesteps are fitted to an external high-step Euler reference, not to the reported FID; headline wins (DiT-XL/2 2.33, SiT-XL/2 2.40, FlowDCN-XL/2 2.35) beat external baselines DPM-Solver++/UniPC. Appendix G's admitted loss-FID mismatch is a correctness risk, not a by-construction reduction.

full rationale

The derivation chain is: (1) Eq. 8 defines the 'optimal solver' as the one minimizing integral error to the continuous ODE; (2) Section 4.2 reduces interpolation to coefficient form (Eq. 16), and Theorems 4.2/4.4/4.5 derive error bounds in Appendices H/I from the paper's own stated assumptions (Assumptions 4.1, 4.3) using standard inequalities — these are in-paper derivations, not results imported from prior work by the authors; (3) Algorithm 2 fits {c, t} by mono-alignment, i.e., MSE/Huber against an L-step Euler reference trajectory, and the headline FID numbers are measured on ImageNet against external solvers (DPM-Solver++ and UniPC) with CFG deliberately excluded from the search. No reported quantity is the fitted parameter renamed: FID is never minimized during search, and the reference trajectory is not the searched solver's own output. Self-citations exist (FlowDCN, Wang et al. 2024b, as search and evaluation model; DDT, Wang et al. 2025, and DMM, Song et al. 2025, in related work), but none is load-bearing: FlowDCN's FID is measured under the same external comparison protocol, and the gains replicate on the externally developed SiT and DiT models. Two flagged gaps weaken the theoretical framing but are not circularity. First, Section 4.2 asserts 'The expectation can be deemed as optimized through massive data and gradient descent' — the identification of gradient-searched coefficients with the Theorem 4.2 optimal expectation-replacement coefficients is asserted rather than proved, so Theorem 4.5's 'theoretically guarantee optimal' conclusion does not formally cover the actual searched solver. Second, Appendix G explicitly concedes 'ablation studies reveal a mismatch between FID performance and Reconstruction error' and promises distribution-matching supervision as future work; the decisive alignment ablation is not run, and Figure 3b shows sensitivity to the reference step count L. These are surrogate-validity and correctness threats to the strength of the FID claims, not circular steps, because the fit target (an Euler-250 trajectory) is not the reported metric. Verdict: no reduction-by-construction; score 1.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The searchable solver is explicitly fitted to reference trajectories: 50,000 generated samples and an L-step Euler target. The coefficients and timesteps are therefore outputs of an optimization, and the headline FID numbers are evaluations of those fitted artifacts. The remaining axioms are analytical assumptions used to argue optimality; several are not justified quantitatively.

free parameters (3)
  • Solver coefficient matrix M = {c_j^i} = Reported per NFE in Appendix J
    Optimized by gradient descent to match a high-step Euler reference trajectory; all reported FID gains are evaluated with these fitted values.
  • Timestep deltas {Delta t_i} = Reported per NFE in Appendix J
    Optimized jointly with coefficients via softmax parameterization; step spacing is part of the searched result, not derived from first principles.
  • Reference trajectory step count L (Euler) = 250, implied by 'Euler-250' in Section 7 and Figure 3b
    Hand-chosen; Figure 3b shows FID improves with larger L, so the headline numbers inherit this choice.
assumptions (5)
  • ad hoc to paper Assumption 4.1: residual of universal interpolation P is O(dx^m)+O(dt^n)
    Stated without derivation in Section 4.2; supports all subsequent error-bound theorems.
  • ad hoc to paper Assumption 4.3: velocity model error is L1-bounded by eta << ||v||
    Used to derive Theorem 4.4; the bound's smallness is asserted, not measured.
  • ad hoc to paper Velocities at different steps are orthogonal with unit norm in the proof of Theorem 4.2
    Appendix H uses ||v_j||^2=1 and orthogonality to show expectation is optimal; not stated in the main text and not true for general diffusion trajectories.
  • ad hoc to paper Interpolation remainder O(dx^m)+O(dt^n) can be omitted because it is much smaller than the other term
    Appendix H and I drop this term with no quantitative estimate.
  • standard math DDPM discrete process corresponds to continuous VP SDE (Song et al. 2020)
    Used in Section 6 to transfer the search to DDPM models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Differentiable Solver Search for Fast Diffusion Sampling." pith.science (2026). https://pith.science/paper/DQ35AC5X

@misc{pith2026250521114,
  author       = {Pith},
  title        = {Pith review of: Differentiable Solver Search for Fast Diffusion Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DQ35AC5X}},
  note         = {Machine review of arXiv:2505.21114}
}
read the original abstract

Diffusion models have demonstrated remarkable generation quality but at the cost of numerous function evaluations. Recently, advanced ODE-based solvers have been developed to mitigate the substantial computational demands of reverse-diffusion solving under limited sampling steps. However, these solvers, heavily inspired by Adams-like multistep methods, rely solely on t-related Lagrange interpolation. We show that t-related Lagrange interpolation is suboptimal for diffusion model and reveal a compact search space comprised of time steps and solver coefficients. Building on our analysis, we propose a novel differentiable solver search algorithm to identify more optimal solver. Equipped with the searched solver, rectified-flow models, e.g., SiT-XL/2 and FlowDCN-XL/2, achieve FID scores of 2.40 and 2.35, respectively, on ImageNet256 with only 10 steps. Meanwhile, DDPM model, DiT-XL/2, reaches a FID score of 2.33 with only 10 steps. Notably, our searched solver outperforms traditional solvers by a significant margin. Moreover, our searched solver demonstrates generality across various model architectures, resolutions, and model sizes.

Figures

Figures reproduced from arXiv: 2505.21114 by the authors.

Figure 1
Figure 1. Visualization of searched Solver Parameters of DDPM/VP and Rectified Flow. We limited the order of solver coefficients of the last two steps for 5/6 NFE. The left images show the absolute value of searched coefficients {c j i }. The right image shows the searched timesteps of different NFE and fitted curves. Abstract Diffusion models have demonstrated remarkable generation quality, but at the cost of numer￾ous funct… view at source ↗
Figure 2
Figure 2. Generated images from Flux.1-dev with Guidance=2.0 and our solver (searched on SiT-XL/2). Euler-Shift3 is the default solver provided by diffusers and Flux community. Our solver(DS-Solver) achieves better visual quality from 5 to 10 steps(NFE). expectation estimation ultimately reduce it to a set of co￾efficients. Inspired by (Xue et al., 2024), we prove that timesteps and these coefficients effectively constitute o… view at source ↗
Figure 3
Figure 3. Ablations studies of Differentiable Solver Search. We evaluate the searched solver on SiT-XL/2, and report the FID performance curve of searched solvers. & Xie, 2023) trained with DDPM scheduling and rectified￾flow models SiT-XL/2 (Ma et al., 2024) and FlowDCN￾XL/2 (Wang et al., 2024b). Our default training setting employs the Lion optimizer (Chen et al., 2024c) with a constant learning rate of 0.01 and no weight de… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The same searched solver on different Rectified-Flow Models. R256 and R512 indicate the generation resolution of given model. We search solver with FlowDCN-B/2 on ImageNet-256 × 256 and evaluate it with SiT-XL/2 and FlowDCN-XL/2 on different resolution datasets. Our se…
Figure 5
Figure 5. Figure 5: The images generated from PixArt-Σ with CFG=2.0 equipped with Our DS-Solver ( searched on DiT-XL/2-R256 ).In comparison to DPM-Solver++ and UniPC, our results consistently exhibit greater clarity and possess more details. Our solver achieves better quality from 5 to 10…
Figure 6
Figure 6. Figure 6: The same searched solver on different Rectified-Flow Models. R256 and R512 indicate the generation resolution of given model. We search solver with FlowDCN-B/2 on ImageNet-256 × 256 and evaluate it with SiT-XL/2 and FlowDCN-XL/2 on different resolution datasets. Our se…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 5 canonical work pages

  1. [1]

    J.1. Solver Searched on SiT-XL/2 NFE TimeDeltas ∆t Coeffcients cj i 5   0.0424 0.1225 0.2144 0.3073 0.3135     0.0 0 .0 0 .0 0 .0 0 .0 −1.17 0 .0 0 .0 0 .0 0 .0 1.07 −1.83 0 .0 0 .0 0 .0 0.0 0 .0 −0.93 0 .0 0 .0 0.0 0 .0 0 .0 −0.71 0 .0   6   0.0389 0.0976 0.161 0.2046 0.2762 0.2217     0.0 0 .0 0 .0 0 .0 0 ...

    work page 1959

  2. [6]

    Consis- tency trajectory models: Learning probability flow ode trajectory of diffusion

    Kim, D., Lai, C.-H., Liao, W.-H., Murata, N., Takida, Y ., Uesaka, T., He, Y ., Mitsufuji, Y ., and Ermon, S. Consis- tency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279,

    arXiv

  3. [7]

    Sdxl-lightning: Progres- sive adversarial diffusion distillation

    Lin, S., Wang, A., and Yang, X. Sdxl-lightning: Progres- sive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929,

    arXiv

  4. [8]

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

    Lipman, Y ., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747,

    arXiv

  5. [9]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003,

    arXiv

  6. [10]

    S., Boffi, N

    Ma, N., Goldstein, M., Albergo, M. S., Boffi, N. M., Vanden- Eijnden, E., and Xie, S. Sit: Exploring flow and diffusion- based generative models with scalable interpolant trans- formers. arXiv preprint arXiv:2401.08740,

    arXiv

  7. [11]

    Align your steps: Op- timizing sampling schedules in diffusion models

    Sabour, A., Fidler, S., and Kreis, K. Align your steps: Op- timizing sampling schedules in diffusion models. arXiv preprint arXiv:2404.14507,

    arXiv

  8. [12]

    and Ho, J

    10 Differentiable Solver Search for Fast Diffusion Sampling Salimans, T. and Ho, J. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512,

    arXiv

Show all 24 references
  1. [13]

    Stylegan-xl: Scaling stylegan to large diverse datasets

    Sauer, A., Schwarz, K., and Geiger, A. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pp. 1–10,

    2022

  2. [14]

    T., Thabet, A., Pumarola, A., and Lipman, Y

    Shaul, N., Perez, J., Chen, R. T., Thabet, A., Pumarola, A., and Lipman, Y . Bespoke solvers for generative flow models. arXiv preprint arXiv:2310.19075,

    arXiv

  3. [15]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. arXiv:2010.02502, October 2020a. URL https://arxiv.org/abs/2010.02502. Song, T., Feng, W., Wang, S., Li, X., Ge, T., Zheng, B., and Wang, L. Dmm: Building a versatile image genera- tion model via distillati...

    2010 arXiv

  4. [17]

    W., Shen, D., Gao, P., Lingelbach, M., Sun, K., Bian, W., Song, G., Liu, Y ., et al

    Wang, F.-Y ., Huang, Z., Bergman, A. W., Shen, D., Gao, P., Lingelbach, M., Sun, K., Bian, W., Song, G., Liu, Y ., et al. Phased consistency model. arXiv preprint arXiv:2405.18407, 2024a. Wang, S., Teng, Y ., and Wang, L. Deep equilibrium object detection. In Proceedings of th...

    arXiv

  5. [18]

    Flowdcn: Exploring dcn-like architectures for fast image generation with arbitrary resolution

    Wang, S., Li, Z., Song, T., Li, X., Ge, T., Zheng, B., and Wang, L. Flowdcn: Exploring dcn-like architectures for fast image generation with arbitrary resolution. arXiv preprint arXiv:2410.22655, 2024b. Wang, S., Tian, Z., Huang, W., and Wang, L. Ddt: Decoupled diffusion trans...

    arXiv

  6. [19]

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

    Zhao, W., Bai, L., Rao, Y ., Zhou, J., and Lu, J. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867,

    arXiv

  7. [20]

    Flowturbo: Towards real-time flow-based image generation with ve- locity refiner

    Zhao, W., Shi, M., Yu, X., Zhou, J., and Lu, J. Flowturbo: Towards real-time flow-based image generation with ve- locity refiner. arXiv preprint arXiv:2409.18128,

    arXiv

  8. [21]

    Trajectory consistency distillation

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

    arXiv

  9. [22]

    To put it simply, under the empowerment of our high-order solver, the performance of DDPM and FM does not differ significantly (8, 9, 10 steps), which contradicts the common belief that FM is stronger at limited sampling steps. F. Text to image Distillation Experiments We unif...

    2017

  10. [23]

    Limitations

    512 0.88s 25 0.9B 20.1 Rectified Flow 512 0.88s 25 0.9B 21.65 Rectified Diffusion 512 0.88s 25 0.9B 21.28 Rectified Flow 512 0.21s 4 0.9B 103.48 PeRFlow 512 0.21s 4 0.9B 22.97 Rectified Diffusion 512 0.21s 4 0.9B 20.64 Ours(Distillation+solver search) 512 0.21s 4 0.9B 18.99 Pe...

    2014

  11. [1883]

    Large scale gan training for high fidelity natural image synthesis

    Brock, A., Donahue, J., and Simonyan, K. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096,

    arXiv

  12. [2020]

    Distilling diffusion models into conditional gans

    Kang, M., Zhang, R., Barnes, C., Paris, S., Kwak, S., Park, J., Shechtman, E., Zhu, J.-Y ., and Park, T. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967,

    arXiv

  13. [2022]

    On the trajectory regularity of ode-based diffusion sampling

    Chen, D., Zhou, Z., Wang, C., Shen, C., and Lyu, S. On the trajectory regularity of ode-based diffusion sampling. arXiv preprint arXiv:2405.11326, 2024a. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y ., Wang, Z., Kwok, J., Luo, P., Lu, H., et al. Pixart- \alpha: Fast train...

    arXiv

  14. [2023]

    Pixart-\sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

    Chen, J., Ge, C., Xie, E., Wu, Y ., Yao, L., Ren, X., Wang, Z., Luo, P., Lu, H., and Li, Z. Pixart-\sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024b. Chen, X., Liang, C., Huang, D., Real, E., Wang, ...

    arXiv

  15. [2024]

    Fast diffusion probabilistic model sampling through the lens of backward error analysis

    Gao, Y ., Pan, Z., Zhou, X., Kang, L., and Chaudhari, P. Fast diffusion probabilistic model sampling through the lens of backward error analysis. arXiv preprint arXiv:2304.11446,

    arXiv

  16. [2025]

    P., Kumar, A., Er- mon, S., and Poole, B

    Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Er- mon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020b. Song, Y ., Dhariwal, P., Chen, M., and Sutskever, I. Consis- tency models. arX...

    2011 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.