Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Improved Immiscible Diffusion: Accelerate Diffusion Training by Reducing Its Miscibility

T0 review · 4 major / 5 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 Separating the noise regions used by different images accelerates diffusion training by more than 4x while preserving diversity, and it can be done with two cheap operations instead of expensive assignment.

desk verdict KNN and image scaling look like real speedups for diffusion training, but the paper does not isolate trajectory miscibility from a plain SNR effect, so the central mechanism outruns the evidence. read the letter →

arxiv 2505.18521 v2 pith:NLFJ63X6 submitted 2025-05-24 cs.CV

Yiheng Li , Feng Liang , Dan Kondratyuk , Masayoshi Tomizuka , Kurt Keutzer , Chenfeng Xu This is my paper · ORCID
classification cs.CV
keywords diffusionmodelstrainingefficiencyimmiscibleKNNnoiseselectionimagescalinggenerativediversityoptimaltransportflowmatching
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 claims that the main bottleneck in diffusion training is 'miscibility': the way different images' forward trajectories overlap and tangle in noise space. It argues that any procedure that keeps each image's noise region separate—not just the expensive linear assignment previously proposed—makes denoising easier and speeds training. Two new cheap procedures are introduced: KNN noise selection, which samples several noises per image and keeps the closest, and image scaling, which multiplies image pixel values so their diffused neighborhoods overlap less. Across consistency models, DDIM, flow matching, and Stable Diffusion on several datasets, the paper reports reaching equivalent generation quality in more than 4x fewer training steps, with no loss of diversity. The paper also claims that denoising is effectively bijective with respect to immiscible diffusion, so confining each image to a noise subregion does not restrict what can be generated.

What carries the argument

The central object is the diffusion trajectory in noise space and its 'miscibility'—the degree to which trajectories from different images overlap. It is measured by the average L2 distance between noise clusters assigned to each image: 0.92 for vanilla DDIM, 4.11 after linear assignment, and 2.17 after KNN. The load-bearing tools are the three implementations that enlarge that separation: batch-wise linear assignment, KNN noise selection (sample k noises and keep the nearest), and image scaling (multiply normalized pixel values by a constant greater than one). The paper uses t-SNE projections of predicted images across denoising layers and per-layer FID to show that separated trajectories give each denoising path a stable goal, so even the noisiest layer can denoise effectively.

What would settle it

Train two otherwise identical diffusion models—one with images scaled to pixel STD 2.0 and one with images at STD 0.5 but with added noise variance reduced by the same factor, so the signal-to-noise ratio matches the scaled run while noise regions stay overlapping; if the second model shows the same training speedup, the gains come from SNR, not miscibility.

Watch

Extended reading notes

Core claim

Reducing the mixing of diffusion trajectories is itself the mechanism that accelerates training, and it can be achieved without image-noise pairing. The paper demonstrates a stable correlation between a noise origin and the generated image: perturbing the noise by 20% leaves the generated image essentially unchanged, which it takes as evidence that vanilla diffusion already assigns each image a local noise region, so immiscible training just makes that assignment explicit. Feature-level analysis shows that the noisiest denoising layers fail to predict the added noise when trajectories are miscible, while immiscible training activates those layers and improves the FID of images predicted at high noise levels. The broadened definition lets the authors treat batch-wise optimal transport, KNN noise selection, and image scaling as instances of one principle, and they report consistent efficiency gains, including a CLIPScore of 28.55 in class-conditional generation matching the baseline, and improved coverage on a robotics pushing task.

Load-bearing premise

The load-bearing premise is that image scaling is a genuine immiscibility intervention rather than a change in the learning problem: multiplying images by a constant greater than one raises the signal-to-noise ratio at every timestep, and the paper does not show that the training speedup survives when the noise level is adjusted to match.

Editorial extensions

If this is right

  • Training diffusion models can be accelerated by a simple per-sample operation at the data level, without changing architecture, loss, or sampling schedule.
  • The bijectivity claim implies that noise-space subregion training does not sacrifice diversity; generation quality at equal steps should improve across unconditional, conditional, fine-tuning, and editing settings.
  • Because KNN costs 0.2 ms per batch of 256 versus 6.7 ms for linear assignment, the approach scales to large batches and high-dimensional latents without O(n^3) overhead.
  • The same principle transfers beyond image generation: in-painting, out-painting, and diffusion-policy robotics planning all show gains, suggesting trajectory miscibility is a general training bottleneck.

Reading between the lines

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

  • If the bijectivity claim holds in full generality, noise space could be partitioned into image-specific cells, opening the door to per-cell caching or curriculum training that vanilla diffusion cannot exploit.
  • The image-scaling experiment does not isolate miscibility, because increasing pixel magnitude raises the signal-to-noise ratio at every timestep; a control that matches SNR while keeping noise regions overlapping would separate the two explanations.
  • The paper's framing suggests that the known speedups from minibatch optimal transport should be attributed to trajectory separation rather than to shorter paths; that reinterpretation could be tested by comparing OT couplings against arbitrary couplings with matched separation.
  • A natural testable extension is to measure per-timestep denoising loss variance under immiscible versus miscible training; the paper's mechanism predicts the variance drop should concentrate in the noisiest layers.
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 / 5 minor

Summary. The paper proposes to broaden the notion of immiscible diffusion from a specific linear assignment between images and noises to any intervention that reduces the mixing (miscibility) of diffusion trajectories from different images. It introduces two new implementations: KNN noise selection (Algorithm 1, §3.3.2) and image scaling (§3.3.3), and reports up to >4x faster training across consistency models, DDIM, flow matching, and Stable Diffusion, on CIFAR-10, ImageNet, and MS-COCO, as well as for image editing and robotics planning. The core claims are that (i) vanilla diffusion already possesses a stable image-noise correlation, so restricting each image to a subregion of noise space preserves diversity; (ii) miscibility reduction eases denoising at noisy layers; and (iii) the training speedups of batch-wise OT are attributable to immiscibility rather than to distance reduction.

Significance. If substantiated, the paper would make a useful practical contribution: the KNN implementation is O(n) per batch and scales far better than the O(n^3) linear assignment of the prior immiscible diffusion work, and the reported consistent speedups across heterogeneous baselines and tasks are encouraging. The authors release code, conduct extensive experiments, and provide feature-level diagnostics (tSNE of predicted images, per-layer FID) that go beyond pure accuracy reporting. However, the central causal claim—that trajectory miscibility, rather than a change in the effective signal-to-noise ratio or in the learning target, causes the acceleration—is not yet convincingly established, because both new implementations alter the denoising task while changing miscibility. The diversity guarantee also rests on qualitative perturbation examples and a prompt-correspondence metric rather than on distributional diversity measures. The practical speedup may well be real, but the mechanism is underdetermined; controlled experiments are needed before the paper can support its conceptual claims.

major comments (4)
  1. [§3.3.2, Algorithm 1] The KNN implementation is not a clean test of miscibility reduction. Conditional on image x, the selected noise n* = argmin_j ||n_j - x|| has a nonzero mean component aligned with x, so E[x_t | x] = sqrt(alpha_t) x + sqrt(1-alpha_t) E[n* | x] contains an extra signal term; this raises the effective SNR at every timestep and can accelerate training independently of trajectory miscibility. The Gaussianity check in §3.3.2 (KL 48.25 vs. 48.60) tests only the marginal distribution of pooled selected noises, which can remain near-Gaussian even when each image's conditional noise is biased toward that image. The paper needs a control that changes trajectory separation while holding conditional denoising difficulty fixed, or an ablation that injects an equivalent SNR boost without KNN selection. Without such a control, the >4x speedup cannot be attributed to the proposed mechanism.
  2. [§3.3.3, §4.6] Image scaling is confounded with both a change in target distribution and a change in signal-to-noise ratio. Multiplying normed images by c>1 increases the L2 separation of diffused-area centers, but it also scales the data and therefore raises the signal-to-noise ratio of every noised sample under a fixed noise schedule. The paper does not state how generated samples are rescaled before FID evaluation, and it does not provide a control with the same SNR increase but no immiscibility intervention. As written, the image-scaling experiment cannot distinguish 'reduced miscibility' from 'an easier denoising task', so it does not support the claim that miscibility reduction is the operative cause.
  3. [§3.1, §4.3] The diversity-preservation claim is not supported by the evidence. Figure 2 is qualitative and based on a small number of hand-picked examples, and the CLIPScore comparison in §4.3 measures image-prompt correspondence, not generation diversity. The paper should report distributional diversity metrics (e.g., FID, recall, or pairwise image distances) for immiscible versus vanilla models, and ideally a quantitative coverage analysis of the noise-to-image mapping, before concluding that the denoising process is effectively bijective with respect to immiscible diffusion.
  4. [§5, OT discussion] The attribution of batch-wise OT's training boost to immiscibility is a central interpretive claim that is not tested. The cited 2% average distance reduction and 4% denoising-STD reduction are summary statistics, not controls; a mechanism can be small on those statistics and still matter through other channels. A direct comparison of linear assignment, KNN, image scaling, and a non-immiscible SNR-matched baseline under identical training budgets is needed to support the claim that trajectory miscibility is the fundamental bottleneck.
minor comments (5)
  1. [§3.2] There are multiple typos, including 'tSN E' and 'stat the average'; please proofread the feature-analysis text and figure captions.
  2. [§3.3.3] The phrase 'the L2 distance between each image is farther' should be rephrased, e.g., 'the pairwise L2 distances between images become larger'.
  3. [Table 4] The header 'Bestk ′s' is malformed; also, the main text should state how k was selected and report sensitivity to k.
  4. [Figure 5] The criterion 'best FID' should be defined precisely (fixed FID threshold vs. minimal achieved FID), since 'training steps required to reach the best FID' is ambiguous when FID curves are non-monotonic.
  5. [Table 1] The in-painting and out-painting FID numbers appear to be single-run evaluations; reporting variance across seeds would strengthen the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the speedup claims are empirical and externally benchmarked; the KNN/image-scaling mechanism has confounds but no definitional reduction.

full rationale

The paper's central claim—that reducing trajectory miscibility accelerates diffusion training—is supported by new implementations (KNN and image scaling) evaluated against external benchmarks (FID, CLIPScore, robotics coverage) and compared with vanilla baselines. The reported speedups are empirical results, not consequences of a definition or of a fitted parameter renamed as a prediction. The KNN algorithm selects the nearest noise by construction, but the measured training-speed improvement is not logically forced by that selection; it is an experimental outcome. Image scaling is labeled 'immiscible' by definition because scaling increases inter-image separation, but the benefit is still an empirical finding; the absence of a control for increased SNR and changed target distribution is a mechanistic confound, not a circular step. Similarly, the KL-divergence check in Sec. 3.3.2 verifies only the marginal noise distribution and does not establish conditional Gaussianity, so the attribution to miscibility is under-supported; this is an omitted control rather than a logical circularity. The self-citation to [17] supplies the original concept and a cited proof, but the present paper's own experiments and external evaluations provide independent content, and no load-bearing argument reduces to the self-citation alone. Therefore no circular step meets the threshold of Eq. X = Eq. Y by construction or a fitted parameter renamed as a prediction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No formal derivation is provided. The argument rests on a geometric definition of miscibility, on the assumption that separating noise clusters is what eases denoising, and on a qualitative perturbation study for diversity. The KNN sample count and image scaling factor are tuned per experiment. No new physical entities are introduced; 'miscibility' is treated as an axiom-loaded concept rather than an invented entity.

free parameters (2)
  • KNN sample count k = k=8 (DDIM CIFAR), k=4 (Flow Matching, Consistency Model CIFAR), k=64 (Stable Diffusion ImageNet)
    Selected per model and dataset from a sweep {1,2,4,8,16,32,64,128}; the central speedup claims use the best k found by this sweep.
  • Image scaling factor (target pixel STD) = 1.0 and 2.0 vs default 0.5
    Chosen by hand; the paper does not specify a selection rule, and the scaling changes the effective signal-to-noise ratio.
assumptions (3)
  • domain assumption Average L2 distance between noise clusters assigned to each image is a valid quantitative measure of trajectory miscibility.
    Used in Section 3.2 to compare vanilla (0.92), linear assignment (4.11), and KNN (2.17); no formal link between this metric and training difficulty is proven.
  • ad hoc to paper Images normed to higher pixel STD make diffusion more immiscible without changing the learning target.
    Sections 3.3.3 and 4.6; the scaling also raises signal-to-noise ratio, and the paper does not show that sampling or evaluation is adapted to avoid a distribution shift.
  • domain assumption Stable generation under 10-20% noise perturbation implies that restricting each image to a noise subregion preserves generative diversity.
    Section 3.1; the evidence is qualitative and does not prove bijectivity or diversity preservation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improved Immiscible Diffusion: Accelerate Diffusion Training by Reducing Its Miscibility." pith.science (2026). https://pith.science/paper/NLFJ63X6

@misc{pith2026250518521,
  author       = {Pith},
  title        = {Pith review of: Improved Immiscible Diffusion: Accelerate Diffusion Training by Reducing Its Miscibility},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NLFJ63X6}},
  note         = {Machine review of arXiv:2505.18521}
}
read the original abstract

The substantial training cost of diffusion models hinders their deployment. Immiscible Diffusion recently showed that reducing diffusion trajectory mixing in the noise space via linear assignment accelerates training by simplifying denoising. To extend immiscible diffusion beyond the inefficient linear assignment under high batch sizes and high dimensions, we refine this concept to a broader miscibility reduction at any layer and by any implementation. Specifically, we empirically demonstrate the bijective nature of the denoising process with respect to immiscible diffusion, ensuring its preservation of generative diversity. Moreover, we provide thorough analysis and show step-by-step how immiscibility eases denoising and improves efficiency. Extending beyond linear assignment, we propose a family of implementations including K-nearest neighbor (KNN) noise selection and image scaling to reduce miscibility, achieving up to >4x faster training across diverse models and tasks including unconditional/conditional generation, image editing, and robotics planning. Furthermore, our analysis of immiscibility offers a novel perspective on how optimal transport (OT) enhances diffusion training. By identifying trajectory miscibility as a fundamental bottleneck, we believe this work establishes a potentially new direction for future research into high-efficiency diffusion training. The code is available at https://github.com/yhli123/Immiscible-Diffusion.

Figures

Figures reproduced from arXiv: 2505.18521 by the authors.

Figure 1
Figure 1. Improved Immiscible Diffusion Theory. (a) While vanilla diffusion trajectories (flows) are mixed (miscible), each noise point is stably correlated to a specific generated image, making many diffusion trajectories irreversible. (b) Those irreversible trajectories would confuse the denoising process. (c) We introduce immiscible diffusion to cut mixed (miscible) diffusion trajectories during training for accelerating d… view at source ↗
Figure 2
Figure 2. Stable correlation between generated images and its noise origins. Here perturbation means another Gaussian noise added to the fixed original Gaussian noise. Note that even with 20% perturbation, images changes are nearly unnoticeable. Only with 30% perturbation, a image object change happens. These demonstrate stable correlation from a noise area to a specific generated image. 3.1 Image-noise Correlation in the Den… view at source ↗
Figure 3
Figure 3. Feature analysis of vanilla (miscible) and immiscible DDIM. Referring to [27], τ = S represents the layer denoising from the pure noise. We show that immiscible diffusion activates the noisiest (τ → S) layers’ denoising functions by clarifying their denoising goals, as shown in the tSNE of denoised images across τ ’s. Such activation results in FID improvements on the denoised images from large τ ’s, which leads to … view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Implementations of Immiscible Diffusion. (a) Miscible Diffusion pairs the batch of images and noises randomly before adding noise to images. (b)(c)(d) Immiscible Diffusion tries to reduce the miscibility of diffusion by (b) L2 linear assignment between the batch of ima…
Figure 5
Figure 5. Figure 5: Immiscible diffusion boosts training efficiency. We show the training steps required to reach the best FID for vanilla models across three diverse diffusion-based architectures. Results consistently show that immiscible diffusion trains significantly faster. 4.1 Experi…
Figure 6
Figure 6. Figure 6: Immiscible diffusion stays effective across diverse image generation tasks. We demon￾strate this with figures showing the training steps v.s. FID of (a)(b) Stable Diffusion class-conditional training and fine-tuning with ImageNet, and (c) conditional training with MS-C…
Figure 7
Figure 7. Figure 7: Immiscible diffusion boosts the performance of diffusion policy. Here the robot (circle) needs to push the T-shaped object (gray) into the desired place (green). Experiment Vanilla (Miscible) Immiscible KNN k=2 Ave. Coverage for Last 10 Ckpts (%) 79.56% 82.83% Max Cove…
Figure 8
Figure 8. Figure 8: Analysis on DDIM with images normed to different pixel value STDs. 5 Discussions In this section, we discuss the distinction of immiscible diffusion to some seemingly similar methods, and compare our immiscible diffusion implementations to highlight their respective ad…
Figure 9
Figure 9. Figure 9: Qualitative Illustration of Immiscible Diffusion in Image-to-image Tasks. We notice that immiscible diffusion can better preserve existed information (i.e. edge information in in-painting and center information in out-painting) so as to provide better overall completed…
Figure 10
Figure 10. Figure 10: tSNE of noise clusters belonging respectively to vanilla, assign and KNN DDIM. A.4 tSNE of Noise Point Clusters for Different Methods A.5 Illustrations Comparing Immiscible Diffusion to Other Common Methods [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Comparison between Diffusion Models, Flow Matching and Immiscible Flow Matching. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 18 canonical work pages

  1. [1]

    Image embedding for denoising generative models.Artificial Intelligence Review, 56(12):14511–14533, 2023

    Andrea Asperti, Davide Evangelista, Samuele Marro, and Fabio Merizzi. Image embedding for denoising generative models.Artificial Intelligence Review, 56(12):14511–14533, 2023

    work page 2023

  2. [2]

    Conditional wasser- stein distances with applications in bayesian ot flow matching, 2024

    Jannis Chemseddine, Paul Hagemann, Gabriele Steidl, and Christian Wald. Conditional wasser- stein distances with applications in bayesian ot flow matching, 2024

    work page 2024

  3. [3]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. InProceedings of Robotics: Science and Systems (RSS), 2023

    2023

  4. [4]

    On implementing 2d rectangular assignment algorithms.IEEE Transactions on Aerospace and Electronic Systems, 52(4):1679–1696, 2016

    David F Crouse. On implementing 2d rectangular assignment algorithms.IEEE Transactions on Aerospace and Electronic Systems, 52(4):1679–1696, 2016

    work page 2016

  5. [5]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. InCVPR09, 2009

    2009

  6. [6]

    Wei Deng, Weijian Luo, Yixin Tan, Marin Biloš, Yu Chen, Yuriy Nevmyvaka, and Ricky T. Q. Chen. Variational schrödinger diffusion models, 2024

    work page 2024

  7. [7]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514–7528, 2021

    2021

  8. [8]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018

    2018

Show all 37 references
  1. [9]

    Blue noise for diffusion models

    Xingchang Huang, Corentin Salaun, Cristina Vasconcelos, Christian Theobalt, Cengiz Oztireli, and Gurprit Singh. Blue noise for diffusion models. InACM SIGGRAPH 2024 Conference Papers, SIGGRAPH ’24, New York, NY , USA, 2024. Association for Computing Machinery

    2024

  2. [10]

    Improving consistency models with generator-induced coupling, 2024

    Thibaut Issenhuth, Ludovic Dos Santos, Jean-Yves Franceschi, and Alain Rakotomamonjy. Improving consistency models with generator-induced coupling, 2024

    2024

  3. [11]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174–24184, 2024

    2024

  4. [12]

    Understanding DDPM latent codes through optimal transport

    Valentin Khrulkov, Gleb Ryzhakov, Andrei Chertkov, and Ivan Oseledets. Understanding DDPM latent codes through optimal transport. InThe Eleventh International Conference on Learning Representations, 2023. 10

    2023

  5. [13]

    Auto-encoding variational bayes, 2022

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022

    2022

  6. [14]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009

    2009

  7. [15]

    Score-based generative modeling secretly minimizes the wasserstein distance

    Dohyun Kwon, Ying Fan, and Kangwook Lee. Score-based generative modeling secretly minimizes the wasserstein distance. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Information Processing Systems, 2022

    2022

  8. [16]

    Minimizing trajectory curvature of ODE-based generative models

    Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ODE-based generative models. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,Proceedings of the 40th International Conference on Mac...

    2023

  9. [17]

    Immiscible diffusion: Accelerating diffusion training with noise assignment

    Yiheng Li, Heyang Jiang, Akio Kodaira, Masayoshi TOMIZUKA, Kurt Keutzer, and Chenfeng Xu. Immiscible diffusion: Accelerating diffusion training with noise assignment. In A. Glober- son, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Ne...

    2024

  10. [18]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InComputer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings,...

    2014

  11. [19]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling, 2023

    2023

  12. [20]

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

    Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  13. [21]

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation

    Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and qiang liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  14. [22]

    Towards a mecha- nistic explanation of diffusion model generalization.arXiv preprint arXiv:2411.19339, 2024

    Matthew Niedoba, Berend Zwartsenberg, Kevin Murphy, and Frank Wood. Towards a mecha- nistic explanation of diffusion model generalization.arXiv preprint arXiv:2411.19339, 2024

    2024 arXiv

  15. [23]

    Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky T. Q. Chen. Multisample flow matching: Straightening flows with minibatch couplings, 2023

    2023

  16. [24]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, June 2022

    2022

  17. [25]

    High-resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022

    2022

  18. [26]

    Laion- 5b: An open large-scale dataset for training next generation image-text models.Advances in neural information processing systems, 35:25278–25294, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion- 5b: An open large-scale dataset for training next generation image-text models.Advances in neural informa...

    2022

  19. [27]

    Denoising diffusion implicit models, 2022

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022

    2022

  20. [28]

    Improved techniques for training consistency models

    Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In The Twelfth International Conference on Learning Representations, 2024

    2024

  21. [29]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, pages 32211–32252. PMLR, 2023. 11

    2023

  22. [30]

    Improving and generalizing flow-based generative models with minibatch optimal transport, 2024

    Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector- Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport, 2024

    2024

  23. [31]

    A closer look at time steps is worthy of triple speed-up for diffusion model training.arXiv preprint arXiv:2405.17403, 2024

    Kai Wang, Mingjia Shi, Yukun Zhou, Zekai Li, Zhihang Yuan, Yuzhang Shang, Xiaojiang Peng, Hanwang Zhang, and Yang You. A closer look at time steps is worthy of triple speed-up for diffusion model training.arXiv preprint arXiv:2405.17403, 2024

    2024 arXiv

  24. [32]

    Patch diffusion: Faster and more data-efficient training of diffusion models.Advances in neural information processing systems, 36:72137–72154, 2023

    Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, Mingyuan Zhou, et al. Patch diffusion: Faster and more data-efficient training of diffusion models.Advances in neural information processing systems, 36:72137–72154, 2023

    2023

  25. [33]

    Exploring straighter trajecto- ries of flow matching with diffusion guidance, 2023

    Siyu Xing, Jie Cao, Huaibo Huang, Xiao-Yu Zhang, and Ran He. Exploring straighter trajecto- ries of flow matching with diffusion guidance, 2023

    2023

  26. [34]

    The emergence of reproducibility and consistency in diffusion models

    Huijie Zhang, Jinfan Zhou, Yifu Lu, Minzhe Guo, Peng Wang, Liyue Shen, and Qing Qu. The emergence of reproducibility and consistency in diffusion models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkam...

    2024

  27. [35]

    Formulating discrete probability flow through optimal transport

    Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  28. [36]

    Non-uniform timestep sampling: Towards faster diffusion model training

    Tianyi Zheng, Cong Geng, Peng-Tao Jiang, Ben Wan, Hao Zhang, Jinwei Chen, Jia Wang, and Bo Li. Non-uniform timestep sampling: Towards faster diffusion model training. InProceedings of the 32nd ACM International Conference on Multimedia, MM ’24, page 7036–7045, New York, NY , U...

    2024

  29. [37]

    There and back again: On the relation between noise and image inversions in diffusion models, 2025

    Łukasz Staniszewski, Łukasz Kuci´nski, and Kamil Deja. There and back again: On the relation between noise and image inversions in diffusion models, 2025. 12 A Technical Appendices and Supplementary Material A.1 Experiment Setup Details Table 3: Image Generation Experiment set...

    2025

Pith tools

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