Contrastive Flow Matching

Ali Farhadi; George Stoica; Judy Hoffman; Ranjay Krishna; Vivek Ramanujan; Xiang Fan

REVIEW 3 major objections 6 minor 1 cited by

Contrastive Flow Matching

T0 review · 3 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 Adding a contrastive repulsion term to conditional flow matching restores flow uniqueness across conditions, yielding up to 9× faster training, 5× fewer denoising steps, and FID reductions of up to 8.9.

desk verdict The paper's own closed-form solution reduces the contrastive term to a global offset, so the claimed per-condition flow separation is not supported by the math. read the letter →

arxiv 2506.05350 v1 pith:H2EGXFBM submitted 2025-06-05 cs.CV

George Stoica , Vivek Ramanujan , Xiang Fan , Ali Farhadi , Ranjay Krishna , Judy Hoffman This is my paper

classification cs.CV

keywords contrastiveflowmatchingconditionalgenerationdiffusionmodelsimagetrainingefficiencyclassifier-freeguidancerepresentationalignment

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

Flow matching trains a generative model by asking a network to predict, at each noisy intermediate state, the straight-line flow that carries noise to a real image. In class-conditional training, these predicted flows can overlap for different classes, which the authors argue produces ambiguous, averaged-looking outputs. The paper claims that adding one contrastive term—for each sample, randomly drawing a second, independent sample and pushing the predicted flow away from that sample's flow—explicitly enforces uniqueness across conditional flows and improves condition separation. On ImageNet-1k with SiT models and on CC3M with an MMDiT text-to-image model, the same architectures trained this way reach a given quality up to 9× faster in training, need up to 5× fewer denoising steps, and lower FID by up to 8.9. The objective is a drop-in addition to any flow-matching loss, adds negligible compute, and stacks with representation alignment and classifier-free guidance.

What carries the argument

The key object is the contrastive flow-matching loss (Eq. 6): a standard conditional flow-matching regression term plus a negative term that pushes the predicted velocity field $v_\theta(x_t,t,y)$ away from the flow of an independently sampled pair $(\tilde{x}, \tilde{\epsilon})$. Because the negative pair is drawn without conditioning on $x_t$ or $y$, its expected flow collapses to a single global vector $\hat{T} = \mathbb{E}[\dot{\alpha}_t \tilde{x} + \dot{\sigma}_t \tilde{\epsilon}]$, so the closed-form optimum is $v^* = (v_{\mathrm{FM}} - \lambda \hat{T})/(1-\lambda)$. This identity carries the argument: it explains the separation effect as a global repulsion from the dataset-mean trajectory, and it yields the adjusted classifier-free-guidance rule $\widehat{\mathrm{CFG}} = (1-\lambda)[w\, v(x_t|y) + (1-w)\, v(x_t|\varnothing)] + \lambda \hat{T}$ used in Section 5.4.

What would settle it

Train the same SiT or MMDiT architectures with the equivalent non-contrastive target that replaces every random negative by the fixed dataset-mean flow $\hat{T}$, i.e. regress $v_\theta$ toward $(\dot{\alpha}_t \hat{x} + \dot{\sigma}_t \epsilon - \lambda \hat{T})/(1-\lambda)$, and compare FID, training speed, and sampling steps with the full ΔFM loss. If the gains disappear, the per-sample identity of negatives carries the signal; if they survive, the claimed contrastive-separation mechanism is equivalent to a constant repulsion and is not what drives the numbers.

Watch

Extended reading notes

Core claim

The paper's central claim is that flow uniqueness across conditions is not a given but something the training objective can enforce. To that end it augments the conditional flow-matching objective with a contrastive term: $L_{\Delta\mathrm{FM}}(\theta) = \mathbb{E}[\, \| v_\theta(x_t,t,y) - (\dot{\alpha}_t \hat{x} + \dot{\sigma}_t \epsilon)\|^2 - \lambda \| v_\theta(x_t,t,y) - (\dot{\alpha}_t \tilde{x} + \dot{\sigma}_t \tilde{\epsilon})\|^2\,]$, where $(\hat{x},\epsilon)$ is the sample being learned and $(\tilde{x},\tilde{\epsilon})$ is an independent random pair drawn from the batch. Minimizing this objective has a closed-form optimum $v^* = (v_{\mathrm{FM}} - \lambda \hat{T})/(1-\lambda)$, with $\hat{T}$ the global mean of all training trajectories, which the paper reads as steering every conditional flow away from the unconditional data trajectory by an amount $\lambda$. Empirically, the authors find that this single term makes class-specific structure appear earlier in the denoising trajectory, improves FID, Inception Score and sFID across SiT-B/2 and SiT-XL/2 on ImageNet at 256 and 512 resolution, improves FID by 5 on CC3M with MMDiT, and remains beneficial when combined with REPA and classifier-free guidance.

Load-bearing premise

The load-bearing premise is that repelling the predicted flow from a randomly drawn negative flow teaches the model to keep each condition's trajectory distinct, even though the network never sees the negative sample's own noisy location and its closed-form effect is only to subtract a global mean trajectory.

Editorial extensions

If this is right

Any conditional flow-matching model can adopt ΔFM by adding one term to the loss, with no extra forward pass, pretrained encoder, or inference-time cost.
At $\lambda=0.05$, ΔFM reaches the baseline's FID with up to 9× fewer training iterations and matches or beats it with up to 5× fewer denoising steps, so the same compute buys better or cheaper generation.
Replacing the flow-matching objective inside REPA with ΔFM improves REPA-SiT FID by up to 6.81, so the contrastive term adds signal beyond representation alignment.
Combined with classifier-free guidance through the corrected rule, ΔFM lowers FID and sFID relative to CFG on the same model, showing the two steering mechanisms are compatible when their conflict is accounted for.
Larger batch sizes increase the benefit, consistent with the negative term behaving as a better Monte Carlo estimate of the global repulsion direction.

Reading between the lines

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

If the closed-form optimum is taken literally, the same performance may be obtainable without random negatives at all, by subtracting an estimated dataset-mean flow from the regression target—a simpler objective the paper does not test.
The early appearance of class-specific structure during denoising suggests ΔFM-trained models are promising starting points for few-step distillation, a regime the paper does not evaluate directly.
Because the repulsion is a single global vector, the method may push rare or visually similar conditions in ways that class imbalance could distort; a long-tail or fine-grained benchmark would test whether this matters.
The same repulsion-from-the-mean mechanism could be realized by decorrelation or orthogonality penalties between conditional velocity fields, offering a family of objectives whose relative merits the paper leaves unexplored.

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.

Desk Editor's Note

read the letter

Here's the short version: the paper's own Appendix B.1 undercuts its central claim. The contrastive term in Eq. (6) has a closed-form optimum that is an affine reparameterization of the flow-matching target. Since the negative sample never enters vθ, the expected effect of the negative term is a constant shift of the regression target—v* = (v_FM − λ T̂)/(1−λ), where T̂ is the mean flow over the dataset. A per-sample contrast whose expected effect is a global offset cannot 'explicitly enforce uniqueness across all conditional flows.' The finite-batch noise around that mean is just gradient noise; it contains no class-dependent signal. So the abstract and intro describe a mechanism the paper's own math says is not happening.

That's not to say the paper is empty. The experimental effort is real: SiT and MMDiT, ImageNet 256/512 and CC3M, multiple metrics, batch-size ablations, and a surprisingly transparent derivation of the closed form. The CFG-coupling analysis in Section 5.4 is a neat piece of algebra. And the reported FID gains, if they replicate, are worth understanding—but the paper never isolates what causes them. The natural baseline is training with the effective target u+ − λ E[u] (or the batch-mean-subtracted target), which is exactly what Eq. (12) says the method is doing. That baseline is absent. Without it, the 9x/5x/8.9-FID numbers cannot be attributed to 'contrastive separation,' and the authors' own derivation suggests they come from something else, like a beneficial rescaling or mean-centering effect.

There are also smaller problems. The novelty claim is overstated: contrastive losses for diffusion/flow training already exist (Contrastive Diffusion Loss, Yu et al., 2024), and it isn't cited. λ=0.05 is fixed after tuning on the target benchmark, and all headline numbers are single runs with no error bars. The CFG comparison in Table 4 is the result of a grid search, not a matched-hyperparameter comparison.

Bottom line: this is a well-executed but mis-claimed paper. The central mechanism is unsupported by the paper's own mathematics; the empirical results are unexplained; and the prior art is thinner than claimed. It could be reframed as a study of how mean subtraction affects flow-matching training, but that work remains to be written. I would not cite it as evidence for contrastive separation. That said, it has enough substance—broad experiments, honest derivation, useful CFG analysis—to warrant peer review rather than a desk reject; a good referee could push the authors toward the right baseline and a much better paper.

Recommendation: send to review, but with a high bar; reject in current form.

Referee Report

3 major / 6 minor

Summary. The paper proposes Contrastive Flow Matching (ΔFM), a modification of the conditional flow-matching loss that appends a negative-flow term to the standard regression objective. The authors claim that this term 'explicitly enforces uniqueness across all conditional flows' and thereby improves condition separation, training speed, sampling efficiency, and FID on ImageNet and CC3M with SiT, SiT+REPA, and MMDiT architectures. The paper includes a closed-form derivation of the minimizer in Appendix B.1, a toy experiment, extensive benchmark tables, and an ablation of the contrastive weight λ. My assessment is that the central theoretical claim is not supported by the paper's own mathematics. Because the network is never evaluated at the negative sample and the negative sample is independent of the conditioning input, the contrastive term reduces to a global affine offset of the standard flow-matching solution. The claimed per-condition flow separation is therefore not a property of the expected loss; the reported empirical gains are left without a valid mechanistic explanation.

Significance. If the claimed mechanism were correct, the paper would make a widely applicable contribution: a training-only loss that improves both training and inference efficiency, is compatible with REPA and CFG, and requires negligible extra compute. The authors also make concrete strengths available: they release code, evaluate on multiple architectures and benchmarks, include an ablation of λ, and provide a closed-form analysis in an appendix. However, the central mechanism is contradicted by the paper's own Appendix B.1. The contrastive term does not create class-dependent behavior in the expected loss; at the optimum it is equivalent to rescaling the flow-matching field and subtracting a global mean velocity. The significance of the empirical results therefore depends on an unsupported reinterpretation of the objective. As written, the paper cannot support its abstract and Section 4 claims.

major comments (3)

[§4, Eq. (6); Appendix B.1, Eqs. (10)–(12)] The central claim that Eq. (6) 'explicitly enforces uniqueness across all conditional flows' is not supported by the objective. The network vθ(xt,t,y) is never evaluated at the negative sample (x̃,ε̃), and the negative sample is independent of (xt,t,y). Minimizing the expected loss pointwise in vθ gives v*(xt,t,y) = (E[α̇t x̂ + σ̇t ε | xt,t,y] − λ E[α̇t x̃ + σ̇t ε̃ | xt,t,y]) / (1−λ). The second conditional expectation is a global, condition-independent constant T̂_t = α̇t E[x̃] + σ̇t E[ε̃] (zero for zero-mean data), so the optimum is v*(xt,t,y) = (v_FM(xt,t,y) − λ T̂_t)/(1−λ). This is an affine rescaling of the standard flow-matching field with no dependence on the identity of the negative sample and no new class-dependent term. The derivation in Eqs. (10)–(12) omits the conditioning on (xt,t,y) and consequently overstates the effect; Eq. (12) also combines a scalar objective with a vector T̂. Consequently, the toy separation in Fig. 3 and the gains in Tables 1–3 cannot be attributed to a contrastive 'separation' mechanism. An alternative mechanism, such as a global norm regularizer, is what the objective actually implements; the paper does not isolate it experimentally.
[§5.4; Appendix B.2] Section 5.4's own description of the closed form contradicts the Abstract. The text states that ΔFM 'yields models which estimate flows away from the data-driven unconditional trajectory' and identifies T̂ as the mean of all sample trajectories from the training set. Subtracting one global mean vector is identical for every condition and cannot separate conditional flows. The CFG coupling in Eqs. (13)–(15) is built on this global-offset interpretation, and the derivation assumes the trained model exactly attains the closed-form optimum, which is not true at finite training. The CFG comparison in Table 4 therefore tests a modified inference-time field rather than the claimed contrastive training mechanism. Before the reported results can be interpreted as evidence for ΔFM, the authors need an experiment that replaces the negative-flow term with a global mean subtraction or a norm penalty and compares FID and trajectory behavior.
[§5.5, Table 5; §5.4, Table 4] The empirical evaluation tunes the key hyperparameters on the same benchmarks used for the headline claims. λ is selected on ImageNet-256 in Table 5 (λ=0.05 gives the best FID on that benchmark) and then fixed for the remaining experiments, and the CFG weights in Table 4 are chosen by grid search on the same ImageNet-256 setting. No validation split is described. Because the reported FID reductions and the 9x/5x speedup claims are measured on the same benchmark family, the results do not establish that ΔFM is generally beneficial independently of this selection. The authors should either choose λ on a held-out validation set or show that the conclusions are robust over a range of λ without selection on the test benchmark.

minor comments (6)

[Table 6] In Table 6, the first two rows (REPA SiT-B/2, batch 256) list exactly the same FID/IS/sFID values as the non-REPA SiT-B/2 rows in Table 1(a), while the caption and surrounding text describe the table as REPA results; this inconsistency must be resolved before the batch-size analysis can be trusted.
[Figure 5] The 9× and 5× speedup claims should be made reproducible: specify the FID threshold used to compare training iterations and denoising steps, report the interpolation method, and improve the legibility of the axes and labels.
[Section 5.5] The text refers to 'Table 5.5' when discussing the λ ablation and the batch-size study; the table numbering should be corrected.
[Abstract and Section 1] There are several copyediting issues: 'guaranteedflows' is missing a space in the Abstract, and the citation list in the Introduction contains a duplicated reference '[11, 11, 20]'.
[Appendix B.1, Eq. (9)] The symbol ∝∼ in Eq. (9) is undefined; please replace it with a standard equivalence or an explicit algebraic step.
[Table 3 and Figure 3] The CC3M comparison in Table 3 reports only FID, and the toy separation in Fig. 3 is qualitative; adding a diversity or condition-alignment metric and a quantitative overlap measure would strengthen these claims.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity; only the ΔFM weight λ is tuned on the reported benchmarks, making the headline FID partly fitted.

fitted input called prediction [Section 5.5 (Table 5.5) and Section 5.1 (Table 2a); reused in Tables 1–3]
"λ= 0.05 is ideal. We show an ablation of the ∆FM weight parameter λ. A too large λ produces degenerate distributions that do not model class structure well. Too low λ is essentially identical to flow-matching, with very little effect on training. λ= 0.05 is best and we use this for all our experiments."

λ is selected by minimizing FID on the same ImageNet-256 REPA SiT-XL/2 setup that is then reported in Table 2a: FID 7.29 appears both as the ablation optimum at λ=0.05 and as the method's headline result. The reported improvement is therefore the best point of an on-benchmark grid rather than an independent prediction. This value is then fixed at 0.05 for all other models and datasets, so the headline numbers in Tables 1–3 inherit a hyperparameter tuned to the target metric. The gain over λ=0 (FM) is a real empirical difference, so this is a mild fitted-input issue rather than full circularity.

full rationale

The paper's derivation chain is not circular: Eq. (6) defines the loss, Appendix B.1 solves it to Eq. (12), and Section B.2 uses Eq. (12) to derive the CFG correction; none of these steps assumes the conclusion. The closed-form solution does reveal that, in expectation, the contrastive term reduces to a global velocity offset independent of y and x_t, which undercuts the paper's claimed per-condition separation mechanism. That is a correctness/falsifiability issue, not a circularity: the claim is contradicted by the paper's own math rather than defined into existence. The only fitted-input concern is the choice of λ=0.05 via an ablation on the same ImageNet-256 REPA SiT-XL/2 benchmark whose FID is later reported, and reuse of this value everywhere. This is a mild methodological weakness, so the circularity score is 2. No load-bearing self-citations are present; REPA [44] and SiT [29] are external prior works with no author overlap with this paper.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the free contrastive weight lambda and on CFG hyperparameters tuned on the target benchmarks. It relies on standard flow-matching interpolant theory and on an unmeasured premise that class flows overlap. No new physical or conceptual entities are introduced; the 'mean trajectory' T is a derived statistic, not an invented entity.

free parameters (3)

lambda (contrastive weight) = 0.05
Chosen by ablation in Table 5 on REPA SiT-XL/2, ImageNet-256; then fixed for all other models and datasets. The reported gains depend on this post-hoc choice.
CFG weight w = 1.85 for DeltaFM, 1.75 for FM
Grid-searched over w in {1.25,1.75,1.8,1.85,2.25} in Table 4; best per model reported.
CFG interval sigma_high = 0.65 for DeltaFM, 0.75 for FM
Grid-searched over {0.50,0.65,0.75,1.0} in Table 4; best per model reported.

assumptions (4)

standard math Stochastic interpolant framework with linear alpha_t and sigma_t (Eq. 1)
Adopted from [2,29] without proof; the derivation relies on this interpolation.
domain assumption Class-conditional distributions overlap in latent flow space, causing ambiguous generations
Motivates the method in Sections 1 and 3; not measured directly in the paper.
domain assumption Batch negatives are representative of the data distribution
Invoked in the batch-size discussion (Section 5.5); contrastive loss behavior depends on this.
standard math The expectation of a squared-loss optimum is the conditional expectation given the network input
The Appendix B.1 derivation should condition on x_t; the paper's unconditional form hides the degeneracy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Contrastive Flow Matching." pith.science (2026). https://pith.science/paper/H2EGXFBM

@misc{pith2026250605350,
  author       = {Pith},
  title        = {Pith review of: Contrastive Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H2EGXFBM}},
  note         = {Machine review of arXiv:2506.05350}
}

read the original abstract

Unconditional flow-matching trains diffusion models to transport samples from a source distribution to a target distribution by enforcing that the flows between sample pairs are unique. However, in conditional settings (e.g., class-conditioned models), this uniqueness is no longer guaranteed--flows from different conditions may overlap, leading to more ambiguous generations. We introduce Contrastive Flow Matching, an extension to the flow matching objective that explicitly enforces uniqueness across all conditional flows, enhancing condition separation. Our approach adds a contrastive objective that maximizes dissimilarities between predicted flows from arbitrary sample pairs. We validate Contrastive Flow Matching by conducting extensive experiments across varying model architectures on both class-conditioned (ImageNet-1k) and text-to-image (CC3M) benchmarks. Notably, we find that training models with Contrastive Flow Matching (1) improves training speed by a factor of up to 9x, (2) requires up to 5x fewer de-noising steps and (3) lowers FID by up to 8.9 compared to training the same models with flow matching. We release our code at: https://github.com/gstoica27/DeltaFM.git.

Figures

Figures reproduced from arXiv: 2506.05350 by the authors.

**Figure 1.** Training with Contrastive Flow-Matching (∆FM) improves natural image generation. (left is baseline, right is with ∆FM) Here we show comparisons between images generated by diffusion models trained on ImageNet-1k (512 × 512). Each pair of images is generated with the same class and initial noise to ensure similar image structure for comparability. We see that our ∆FM objective encourages significantly more coherent i… view at source ↗

**Figure 2.** ∆FM yields more discriminative and higher quality trajectories. (left) shows the result of standard flow-matching, where flows are straight but end up overlapping for similar class distributions. (right) shows how the addition of the ∆FM objective results in more distinct flows, resulting in images which are more representative of their respective classes. structure the representations at an intermediate layer with … view at source ↗

**Figure 3.** Contrastive Flow-Matching intrinsically separates flows between classes. We train a small three layer MLP flowmatching model to transport between a two dimensional multivariate noise distribution (violet) and two independent blue and orange class distributions respectively. The class distributions are designed to have ∼ 50% overlap, and we plot the learned class-conditioned flows between noise samples and each res… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Contrastive flow-matching (∆FM) denoises significantly more efficiently than flow-matching. We visualize the expected final image estimated by a flow-model when denoised every 5 steps for trajectories of length 30 steps using the SDE Euler-Maruyama sampler and do not u…

**Figure 5.** Figure 5: ∆FM requires significantly fewer training iterations and inference-time denoising steps. We plot FID-50k on ImageNet 256x256 with different numbers of training iterations and denoising steps. We see that ∆FM outperforms the baseline with 9× fewer training iterations a…

**Figure 6.** Figure 6: CC3M side-by-side generations between a REPA-MMDiT model trained with flow-matching (left) and ∆FM (right). Models are trained for 400K iterations using a batch-size of 256 and images are generated without classifier-free guidance and using NFE=50. 13 [PITH_FULL_IMAGE…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

MixFlow Training: Alleviating Exposure Bias with Slowed Interpolation Mixture
cs.CV 2025-12 conditional novelty 6.0 of 10

Training diffusion models on a mixture of higher-noise interpolations (MixFlow) improves generation FID across SiT, REPA, RAE and SD3.5, reaching ImageNet 256 gFID 1.43 after post-training.

Reference graph

Works this paper leans on

46 extracted references · 26 canonical work pages · cited by 1 Pith paper

[1]

Building nor- malizing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571, 2022

Michael S Albergo and Eric Vanden-Eijnden. Building nor- malizing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571, 2022. 3

arXiv 2022
[2]

Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023

Michael S Albergo, Nicholas M Boffi, and Eric Vanden- Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023. 3

arXiv 2023
[3]

All are worth words: A vit backbone for diffusion models

Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. InCVPR, 2023. 5

work page 2023
[4]

Tripletgan: Training generative model with triplet loss, 2017

Gongze Cao, Yezhou Yang, Jie Lei, Cheng Jin, Yang Liu, and Mingli Song. Tripletgan: Training generative model with triplet loss, 2017. 3

work page 2017
[5]

Emerg- ing properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. InICCV,
[6]

Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations. In Advances in Neural Information Processing Systems, 2018. 3

work page 2018
[7]

A simple framework for contrastive learning of visual representations.ICLR, 2020

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geof- frey Hinton. A simple framework for contrastive learning of visual representations.ICLR, 2020. 8

work page 2020
[8]

Diffusion posterior sam- pling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022

Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sam- pling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022. 3

arXiv 2022

Show all 46 references

[9]

Cfg++: Manifold-constrained clas- sifier free guidance for diffusion models.arXiv preprint arXiv:2406.08070, 2024

Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained clas- sifier free guidance for diffusion models.arXiv preprint arXiv:2406.08070, 2024. 3

2024 arXiv
[10]

Imagenet: A large-scale hierarchical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 5, 7

2009
[11]

Improving image syn- thesis with diffusion-negative sampling, 2024

Alakh Desai and Nuno Vasconcelos. Improving image syn- thesis with diffusion-negative sampling, 2024. 2

2024
[12]

Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021

Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021. 3, 5

2021
[13]

An image is worth 16x16 words: Transformers for image recognition at scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

2021
[14]

Scaling rectified flow transformers for high-resolution image synthesis.ICML, 2024

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim En- tezari, Jonas M¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synt...

2024
[15]

Momentum contrast for unsupervised visual repre- sentation learning.CVPR, 2020

Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual repre- sentation learning.CVPR, 2020. 8

2020
[16]

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

Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bern- hard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 9

2017
[17]

Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. 3

2022 arXiv
[18]

Classifier-free diffusion guidance, 2022

Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 2

2022
[19]

Denoising diffu- sion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. InAdvances in Neural Information Processing Systems, 2020. 3

2020
[20]

Guiding a dif- fusion model with a bad version of itself.arXiv preprint arXiv:2406.02507, 2024

Tero Karras, Miika Aittala, Tuomas Kynk¨a¨anniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a dif- fusion model with a bad version of itself.arXiv preprint arXiv:2406.02507, 2024. 2, 3

2024 arXiv
[21]

Dynamic negative guidance of diffusion models: Towards immediate content removal

Felix Koulischer, Johannes Deleu, Gabriel Raya, Thomas De- meester, and Luca Ambrogioni. Dynamic negative guidance of diffusion models: Towards immediate content removal. In Neurips Safe Generative AI Workshop 2024. 3

2024
[22]

Improved precision and recall metric for assessing generative models.NeurIPS, 2019

Tuomas Kynk¨a¨anniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models.NeurIPS, 2019. 5

2019
[23]

Applying guidance in a limited interval improves sample and distribution quality in diffusion models.arXiv preprint arXiv:2404.07724, 2024

Tuomas Kynk¨a¨anniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models.arXiv preprint arXiv:2404.07724, 2024. 3

2024 arXiv
[24]

Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matthew Le. Flow matching for generative modeling. InICLR, 2023. 3

2023
[25]

Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matthew Le. Flow matching for generative modeling. InICLR, 2023. 1, 3, 4

2023
[26]

Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022

Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. 3

2022
[27]

Contrastive energy prediction for exact energy-guided diffusion sampling in offline reinforcement learning

Cheng Lu, Huayu Chen, Jianfei Chen, Hang Su, Chongxuan Li, and Jun Zhu. Contrastive energy prediction for exact energy-guided diffusion sampling in offline reinforcement learning. InInternational Conference on Machine Learning, pages 22825–22855. PMLR, 2023. 3

2023
[28]

Latent consistency models: Synthesizing high- resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023

Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high- resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023. 3

2023 arXiv
[29]

Albergo, Nicholas M

Nanye Ma, Mark Goldstein, Michael S. Albergo, Nicholas M. Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. 2024. 1, 2, 3, 5, 6

2024
[30]

Generating images with sparse representations

Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W Battaglia. Generating images with sparse representations. arXiv preprint arXiv:2103.03841, 2021. 5

2021 arXiv
[31]

Maxime Oquab, Timoth´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael R...

2024
[32]

High-resolution image synthesis with latent diffusion models

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 5

2022
[33]

Berg, and Li Fei-Fei

Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge,
[34]

Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022

Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022. 3

2022 arXiv
[35]

Improved techniques for training gans.Advances in neural information processing systems, 29, 2016

Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans.Advances in neural information processing systems, 29, 2016. 5

2016
[36]

Facenet: A unified embedding for face recognition and clus- tering

Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clus- tering. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 815823. IEEE, 2015. 3

2015
[37]

Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning

Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. InProceed- ings of ACL, 2018. 2, 5, 6

2018
[38]

Gradient-free classifier guidance for diffusion model sam- pling.arXiv preprint arXiv:2411.15393, 2024

Rahul Shenoy, Zhihong Pan, Kaushik Balakrishnan, Qisen Cheng, Yongmoon Jeon, Heejune Yang, and Jaewon Kim. Gradient-free classifier guidance for diffusion model sam- pling.arXiv preprint arXiv:2411.15393, 2024. 3

2024 arXiv
[39]

Denoising diffusion implicit models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021. 3

2021
[40]

Loss-guided diffusion models for plug-and-play con- trollable generation

Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mar- dani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play con- trollable generation. InInternational Conference on Machine Learning, pages 32483–32498. PMLR, 2023. 3

2023
[41]

Score-based generative modeling in latent space, 2021

Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space, 2021. 3

2021
[42]

Im- proved distribution matching distillation for fast image syn- thesis.arXiv preprint arXiv:2405.14867, 2024

Tianwei Yin, Micha¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Im- proved distribution matching distillation for fast image syn- thesis.arXiv preprint arXiv:2405.14867, 2024. 3

2024 arXiv
[43]

One-step diffusion with distribution matching distillation

Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 6613–6623...

2024
[44]

Representa- tion alignment for generation: Training diffusion transformers is easier than you think, 2024

Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representa- tion alignment for generation: Training diffusion transformers is easier than you think, 2024. 1, 2, 3, 5, 6, 7, 8

2024
[45]

Egsde: Un- paired image-to-image translation via energy-guided stochas- tic differential equations.Advances in Neural Information Processing Systems, 35:3609–3623, 2022

Min Zhao, Fan Bao, Chongxuan Li, and Jun Zhu. Egsde: Un- paired image-to-image translation via energy-guided stochas- tic differential equations.Advances in Neural Information Processing Systems, 35:3609–3623, 2022. 3 10

2022
[46]

Long and short guidance in score identity distilla- tion for one-step text-to-image generation.arXiv preprint arXiv:2406.01561, 2024

Mingyuan Zhou, Zhendong Wang, Huangjie Zheng, and Hai Huang. Long and short guidance in score identity distilla- tion for one-step text-to-image generation.arXiv preprint arXiv:2406.01561, 2024. 3 11 A. Text-to-Image Qualitative Results We visualize generations between our REP...

2024 arXiv

Pith tools

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

[1] [1]

Building nor- malizing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571, 2022

Michael S Albergo and Eric Vanden-Eijnden. Building nor- malizing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571, 2022. 3

arXiv 2022

[2] [2]

Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023

Michael S Albergo, Nicholas M Boffi, and Eric Vanden- Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023. 3

arXiv 2023

[3] [3]

All are worth words: A vit backbone for diffusion models

Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. InCVPR, 2023. 5

work page 2023

[4] [4]

Tripletgan: Training generative model with triplet loss, 2017

Gongze Cao, Yezhou Yang, Jie Lei, Cheng Jin, Yang Liu, and Mingli Song. Tripletgan: Training generative model with triplet loss, 2017. 3

work page 2017

[5] [5]

Emerg- ing properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. InICCV,

[6] [6]

Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations. In Advances in Neural Information Processing Systems, 2018. 3

work page 2018

[7] [7]

A simple framework for contrastive learning of visual representations.ICLR, 2020

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geof- frey Hinton. A simple framework for contrastive learning of visual representations.ICLR, 2020. 8

work page 2020

[8] [8]

Diffusion posterior sam- pling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022

Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sam- pling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022. 3

arXiv 2022