Continuous Adversarial MeanFlow Transfer

Yara Bahram; Zahra Dehghani; Mélodie Desbos; Eric Granger; Pablo Piantanida; Mohammadhadi Shateri

cs.LGAug 20, 2026

Continuous Adversarial MeanFlow Transfer

Authors: Yara Bahram, Zahra Dehghani, Mélodie Desbos, Eric Granger, Pablo Piantanida, Mohammadhadi Shateri

Organizations: LIVIA · ILLS · ÉTS Montreal · Mila – Quebec AI Institute ·

Abstract

Training fast generators on new domains with limited data remains challenging for two reasons. First, adapting a pretrained diffusion or flow model to a new domain leaves its costly multi-step sampling unaddressed, and existing acceleration methods are tied to the source parameterization-- $ε$ , $x$ , $v$ , or $u$ --leaving heterogeneous pretrained models with no common acceleration target. Second, while adversarial refinement is proven effective for few-step quality, it is formulated only for instantaneous-velocity flows, not for the finite-interval average velocities that MeanFlow (MF) models predict. We address both problems. We propose MeanFlow-Transfer, which maps heterogeneous source outputs into a shared velocity representation, uses it to initialize an MF generator from the source weights, and optimizes an MF objective on the target domain. This unifies adaptation and acceleration in a single training loop across a broad range of pretrained models. We then introduce Continuous Adversarial MeanFlow, a post-training stage that extends continuous adversarial flow models from instantaneous velocities to MF's finite-interval average velocities. CAMF contrasts changes in a learned potential between real and predicted interval endpoints, recovering fine detail that MF regression averages away, and reduces to the instantaneous criterion in the vanishing-interval limit. Adapting four ImageNet-based source models--DiT ( $ε$ ), SiT ( $v$ ), JiT ( $x$ ), iMF ( $u$ )--to five target domains, MF-T with CAMF matches or exceeds the fine-tuned teacher in FID and FDD at up to $125\times$ fewer Neural Function Evaluations (NFEs), while CAMF improves MF-T's few-step FID by $29\%$ on average.

Figures & tables

Figure 1: MF-T + CAMF adapts a source-domain diffusion or flow model with any output parameterization into a high-quality few-step MeanFlow (MF) generator in a target domain with $\text{NFE}_{\text{trg}}\mathop{\ll}\text{NFE}_{\text{src}}$ ( $\text{NFE}_{\text{trg}}\sim\text{NFE}_{\text{src}}$ if source model is MF). Source and target domains: ImageNet ( 8 ) and Birds ( 43 ) .

Panel of Figure 2 — Figure 2: Target-domain generated images using MF-T+CAMF initialized from different source models. Images are uncurated.

Panel of Figure 3 — Figure 3: Overview of our two-stage framework. (a) MF-T : a pretrained source generator with an arbitrary output parameterization is mapped to a shared velocity representation and adapted into a target-domain MeanFlow (MF) model that predicts the finite-interval average velocity $\bm{u_{\theta}}$ . (b) CAMF : the MF model transports $z_{t}$ to the predicted endpoint ${\color[rgb]{0.8711,0.7266,0.1094}\bm{\widehat{z}_{r}}}=z_{t}-(t-r){\color[rgb]{0.8711,0.7266,0.1094}\bm{u_{\theta}}}$ , and the discriminator compares the average change of the potential $D_{\psi}$ along the real and model-predicted target-domain transport segments, through $\mathcal{A}_{\mathrm{real}}$ and $\mathcal{A}_{\mathrm{fake}}$ .

Source prediction space
$\bm{x}$	$\bm{\epsilon}$	$\bm{v}$	$\bm{u}$
Source prediction
$x_{\phi}(z_{t})$	$\epsilon_{\phi}(z_{t})$	$v_{\phi}(z_{t})$	$u_{\phi}(z_{t})$
Recovered instantaneous velocity $v_{\psi}(z_{t})$
$\dfrac{z_{t}-x_{\phi}(z_{t})}{t}$	$\dfrac{\epsilon_{\phi}(z_{t})-z_{t}}{1-t}$	$v_{\phi}(z_{t})$	$u_{\phi}(z_{t},t,t)^{\dagger}$
MeanFlow-Transfer initialization

Table 1: MF-T initialization from any source parameterization. Under the linear interpolation path, MF-T recovers a consistent instantaneous velocity from each prediction space and uses it to initialize the target MeanFlow model

u_{\theta}

FID

\downarrow

FDD

\downarrow

ImageNet Init.

Method

NFE

ArtB.

Calt.

CUB

Food

Cars

Avg.

ArtB.

Calt.

CUB

Food

Cars

Avg.

AFM

4

{\color[rgb]{0.55,0.55,0.55}15.84}

{\color[rgb]{0.55,0.55,0.55}55.97}

{\color[rgb]{0.55,0.55,0.55}9.86}

{\color[rgb]{0.55,0.55,0.55}11.86}

{\color[rgb]{0.55,0.55,0.55}12.39}

21.18

{\color[rgb]{0.55,0.55,0.55}290.57}

{\color[rgb]{0.55,0.55,0.55}1007.22}

{\color[rgb]{0.55,0.55,0.55}570.89}

{\color[rgb]{0.55,0.55,0.55}476.94}

{\color[rgb]{0.55,0.55,0.55}440.94}

557.31

CAMF

4

{\color[rgb]{0.55,0.55,0.55}10.80}

{\color[rgb]{0.55,0.55,0.55}44.42}

{\color[rgb]{0.55,0.55,0.55}9.34}

{\color[rgb]{0.55,0.55,0.55}10.60}

{\color[rgb]{0.55,0.55,0.55}11.19}

17.27

{\color[rgb]{0.55,0.55,0.55}221.40}

{\color[rgb]{0.55,0.55,0.55}812.31}

{\color[rgb]{0.55,0.55,0.55}367.57}

{\color[rgb]{0.55,0.55,0.55}409.74}

{\color[rgb]{0.55,0.55,0.55}293.50}

420.90

MF-T

4

{\color[rgb]{0.55,0.55,0.55}10.60}

{\color[rgb]{0.55,0.55,0.55}27.54}

{\color[rgb]{0.55,0.55,0.55}7.26}

{\color[rgb]{0.55,0.55,0.55}10.58}

{\color[rgb]{0.55,0.55,0.55}9.69}

13.13

{\color[rgb]{0.55,0.55,0.55}273.28}

{\color[rgb]{0.55,0.55,0.55}471.09}

{\color[rgb]{0.55,0.55,0.55}310.62}

{\color[rgb]{0.55,0.55,0.55}465.38}

{\color[rgb]{0.55,0.55,0.55}276.53}

359.38

MF-T + AFM

4

{\color[rgb]{0.55,0.55,0.55}8.67}

{\color[rgb]{0.55,0.55,0.55}22.19}

{\color[rgb]{0.55,0.55,0.55}3.01}

{\color[rgb]{0.55,0.55,0.55}7.03}

{\color[rgb]{0.55,0.55,0.55}4.13}

9.01

{\color[rgb]{0.55,0.55,0.55}180.29}

{\color[rgb]{0.55,0.55,0.55}371.61}

{\color[rgb]{0.55,0.55,0.55}75.72}

{\color[rgb]{0.55,0.55,0.55}298.54}

{\color[rgb]{0.55,0.55,0.55}141.25}

213.48

MF-T + CAMF

4

{\color[rgb]{0.55,0.55,0.55}6.71}

{\color[rgb]{0.55,0.55,0.55}22.10}

{\color[rgb]{0.55,0.55,0.55}2.88}

{\color[rgb]{0.55,0.55,0.55}4.65}

{\color[rgb]{0.55,0.55,0.55}3.07}

\underline{\mathbf{7.88}}

{\color[rgb]{0.55,0.55,0.55}146.40}

{\color[rgb]{0.55,0.55,0.55}378.09}

{\color[rgb]{0.55,0.55,0.55}84.24}

{\color[rgb]{0.55,0.55,0.55}239.28}

{\color[rgb]{0.55,0.55,0.55}102.37}

\underline{\mathbf{190.08}}

Table 2: Quantitative comparison with baselines across four ImageNet-initialized models and five target datasets. We report FID and FDD across different NFEs after adaptation to target-domain. AFM and CAMF as stand-alone methods are only tested on iMF/XL-2 where the initial model is MF-based (with MF-T

\approx

FT on iMF). Bold : best average result under each model initialization and NFE. Underlined : best average result under each model initialization across all NFEs.

Panel of Figure 4 — Figure 4: Comparison of DDPM- $\bm{\epsilon}$ , DDPM- $\bm{v}$ , and Transport- $\bm{v}$ for $\text{DiT}\rightarrow\text{SiT}$ and $\text{DiT}\rightarrow\texttt{MF-T}$ transformation under domain shift. Source domain and model are ImageNet and DiT/XL-2.

	FID $\downarrow$			FDD $\downarrow$
Centering penalty $\mathcal{L}_{\mathrm{cp}}$	ArtB.	CUB	Cars	ArtB.	CUB	Cars
$\varnothing$	${\color[rgb]{0.55,0.55,0.55}7.49}$	${\color[rgb]{0.55,0.55,0.55}2.85}$	${\color[rgb]{0.55,0.55,0.55}3.13}$	${\color[rgb]{0.55,0.55,0.55}160.24}$	${\color[rgb]{0.55,0.55,0.55}86.10}$	${\color[rgb]{0.55,0.55,0.55}111.71}$
$D_{z_{t}}^{2}$	${\color[rgb]{0.55,0.55,0.55}7.81}$	${\color[rgb]{0.55,0.55,0.55}2.79}$	${\color[rgb]{0.55,0.55,0.55}3.24}$	${\color[rgb]{0.55,0.55,0.55}157.70}$	${\color[rgb]{0.55,0.55,0.55}82.70}$	${\color[rgb]{0.55,0.55,0.55}125.14}$
$D_{z_{t}}^{2}+D_{\widehat{z}_{r}}^{2}$	${\color[rgb]{0.55,0.55,0.55}6.88}$	${\color[rgb]{0.55,0.55,0.55}2.80}$	${\color[rgb]{0.55,0.55,0.55}3.36}$	${\color[rgb]{0.55,0.55,0.55}151.95}$	${\color[rgb]{0.55,0.55,0.55}84.60}$	${\color[rgb]{0.55,0.55,0.55}128.25}$
$D_{z_{t}}^{2}+D_{z_{r}}^{2}$	${\color[rgb]{0.55,0.55,0.55}8.11}$	${\color[rgb]{0.55,0.55,0.55}2.85}$	${\color[rgb]{0.55,0.55,0.55}3.24}$	${\color[rgb]{0.55,0.55,0.55}170.76}$	${\color[rgb]{0.55,0.55,0.55}83.89}$	${\color[rgb]{0.55,0.55,0.55}104.18}$
$(D_{z_{t}}+D_{\hat{z}_{r}})^{2}$	${\color[rgb]{0.55,0.55,0.55}7.68}$	${\color[rgb]{0.55,0.55,0.55}2.77}$	${\color[rgb]{0.55,0.55,0.55}3.27}$	${\color[rgb]{0.55,0.55,0.55}159.88}$	${\color[rgb]{0.55,0.55,0.55}82.37}$	${\color[rgb]{0.55,0.55,0.55}123.60}$

Table 3:

\mathcal{L}_{\text{cp}}

ablation on iMF/XL-2 MF-T + CAMF (

\text{NFE}=4

, best result within

30\text{K}+30\text{K}

training steps,

D_{z_{t}}:D_{\psi}(z_{t})

D_{z_{r}}:D_{\psi}(z_{r})

D_{\widehat{z}_{r}}:D_{\psi}(\widehat{z}_{r})

). Rows are ordered from least to most complex centering penalty.

		FID $\downarrow$			FDD $\downarrow$
Method	$\mathcal{\lambda}_{\mathrm{ot}}$	ArtB.	Calt.	CUB	ArtB.	Calt.	CUB
CAMF	$4$	${\color[rgb]{0.55,0.55,0.55}12.18}$	${\color[rgb]{0.55,0.55,0.55}59.80}$	${\color[rgb]{0.55,0.55,0.55}12.52}$	${\color[rgb]{0.55,0.55,0.55}231.13}$	${\color[rgb]{0.55,0.55,0.55}974.38}$	${\color[rgb]{0.55,0.55,0.55}454.71}$
CAMF	$0$	$\mathbf{10.80}$	$\mathbf{44.42}$	$\mathbf{9.34}$	$\mathbf{221.40}$	$\mathbf{812.31}$	$\mathbf{367.57}$
MF-T + CAMF	$4$	${\color[rgb]{0.55,0.55,0.55}7.59}$	${\color[rgb]{0.55,0.55,0.55}24.02}$	${\color[rgb]{0.55,0.55,0.55}3.23}$	${\color[rgb]{0.55,0.55,0.55}159.11}$	${\color[rgb]{0.55,0.55,0.55}408.27}$	${\color[rgb]{0.55,0.55,0.55}98.86}$
MF-T + CAMF	$0$	$\mathbf{6.71}$	$\mathbf{22.10}$	$\mathbf{2.88}$	$\mathbf{146.4}$	$\mathbf{378.09}$	$\mathbf{84.24}$

Table 4:

\mathcal{L}_{\text{ot}}

ablation on iMF/XL-2 (

\text{NFE}=4

, best result within

30\text{K}+30\text{K}

training steps for MF-T + CAMF and

30\text{K}

for CAMF ).

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

		Training			Inference
Family	Stage	Updates	Batch	GPU-h	NFE	TFLOPs/img	Latency (ms)
iMF	MF-T	$40\text{k}$	$32$	$12.85$	$4$	$0.337$	$17.48$
iMF	$+$ CAMF	$30\text{k}$	$2$	$12.88$	$4$	$0.337$	$17.48$
SiT	FT (baseline)	$30\text{k}$	$32$	$1.54$	$500$	$33.33$	$735.0$
	MF-T	$30\text{k}$	$32$	$3.32$	$4$	$0.267$	$5.88$
	$+$ CAMF	$30\text{k}$	$2$	$6.99$	$4$	$0.267$	$5.88$

Appendix

Table 5: Computational cost on CUB-200 , single NVIDIA H100 80GB SXM. “Updates” counts generator optimizer steps. GPU-hours include periodic in-training evaluation (once each

2.5

K training steps generating

10

K samples for four-step evaluation). Latency is the forward sampling time at batch

32

, warmup-then-timed and amortized per image.

			IS $\uparrow$
ImageNet Init.	Method	NFE	ArtB.	Calt.	CUB	Food	Cars	Avg.
	AFM	$4$	${\color[rgb]{0.55,0.55,0.55}6.98}$	${\color[rgb]{0.55,0.55,0.55}17.27}$	${\color[rgb]{0.55,0.55,0.55}5.14}$	${\color[rgb]{0.55,0.55,0.55}4.16}$	${\color[rgb]{0.55,0.55,0.55}3.33}$	$7.38$
	CAMF	$4$	${\color[rgb]{0.55,0.55,0.55}7.94}$	${\color[rgb]{0.55,0.55,0.55}23.54}$	${\color[rgb]{0.55,0.55,0.55}6.16}$	${\color[rgb]{0.55,0.55,0.55}5.28}$	${\color[rgb]{0.55,0.55,0.55}3.66}$	$9.32$
	MF-T	$4$	${\color[rgb]{0.55,0.55,0.55}7.36}$	${\color[rgb]{0.55,0.55,0.55}31.46}$	${\color[rgb]{0.55,0.55,0.55}6.08}$	${\color[rgb]{0.55,0.55,0.55}6.37}$	${\color[rgb]{0.55,0.55,0.55}3.41}$	$10.94$
	MF-T + AFM	$4$	${\color[rgb]{0.55,0.55,0.55}7.62}$	${\color[rgb]{0.55,0.55,0.55}36.69}$	${\color[rgb]{0.55,0.55,0.55}5.67}$	${\color[rgb]{0.55,0.55,0.55}5.47}$	${\color[rgb]{0.55,0.55,0.55}3.36}$	$11.76$
	MF-T + CAMF	$4$	${\color[rgb]{0.55,0.55,0.55}8.03}$	${\color[rgb]{0.55,0.55,0.55}36.16}$	${\color[rgb]{0.55,0.55,0.55}5.69}$	${\color[rgb]{0.55,0.55,0.55}5.76}$	${\color[rgb]{0.55,0.55,0.55}3.23}$	$\underline{\mathbf{11.77}}$

Appendix

Table 6: Quantitative comparison with baselines across four ImageNet-initialized models and five target datasets. We report Inception Score (IS) across different NFEs after adaptation to target-domain. AFM and CAMF as stand-alone methods are only tested on iMF/XL-2 where the initial model is MF-based (with MF-T

\approx

FT on iMF). Bold : best average result under each model initialization and NFE. Underlined : best average result under each model initialization across all relevant NFEs.

	iMF-XL/2	SiT-XL/2	DiT-XL/2	JiT-H/16
	Source model
Parameters	610M	675M + 49M	675M + 49M	953M
Reference NFE	$1/2$	$250\times 2$	$250\times 2$	$50\times 2$
Reference FID	$1.72/1.54$	2.06	2.27	1.86
Prediction space	$u$	$v$	$\epsilon$	$x$
Loss space	$v$	$v$	$\epsilon$	$v$

Appendix

Table 7: Backbone configurations and hyperparameters. All models are ImageNet-based and utilize a

256\times 256

resolution. Reference NFE for SiT and DiT is

250\times 2

because the standard reported results use 250 denoising steps with classifier-free guidance. Batch sizes are effective (per-device batch

\times

gradient accumulation).

Panel of Figure 5 — Figure 5: Target-domain generated images using MF-T+CAMF initialized from different source models. Images are uncurated.

Diffusion Model Adaptation

See all 25 papers

Apr 13, 2026

Continuous Adversarial Flow Models

We propose continuous adversarial flow models, a type of continuous-time flow model trained with an adversarial objective. Unlike flow matching, which uses a fixed mean-squared-error criterion, our approach introduces a learned discriminator to guide training. This change in objective induces a different generalized distribution, which empirically produces samples that are better aligned with the target data distribution. Our method is primarily proposed for post-training existing flow-matching models, although it can also train models from scratch. On the ImageNet 256px generation task, our post-training substantially improves the guidance-free FID of latent-space SiT from 8.26 to 3.63 and of pixel-space JiT from 7.17 to 3.57. It also improves guided generation, reducing FID from 2.06 to 1.53 for SiT and from 1.86 to 1.80 for JiT. We further evaluate our approach on text-to-image generation, where it achieves improved results on both the GenEval and DPG benchmarks.

Shanchuan Lin, Ceyuan Yang, Zhijie Lin +2

ByteDance Seed

Flow-Based Generative Modeling Image Generation

Ask PDF

Apr 10, 2026

Discrete Meanflow Training Curriculum

Flow-based image generative models exhibit stable training and produce high quality samples when using multi-step sampling procedures. One-step generative models can produce high quality image samples but can be difficult to optimize as they often exhibit unstable training dynamics. Meanflow models exhibit excellent few-step sampling performance and tantalizing one-step sampling performance. Notably, MeanFlow models that achieve this have required extremely large training budgets. We significantly decrease the amount of computation and data budget it takes to train Meanflow models by noting and exploiting a particular discretization of the Meanflow objective that yields a consistency property which we formulate into a ``Discrete Meanflow'' (DMF) Training Curriculum. Initialized with a pretrained Flow Model, DMF curriculum reaches one-step FID 3.36 on CIFAR-10 in only 2000 epochs. We anticipate that faster training curriculums of Meanflow models, specifically those fine-tuned from existing Flow Models, drives efficient training methods of future one-step examples.

Chia-Hong Hsu, Frank Wood

Department of Computer Science, University of British Columbia, Vancouver, BC, Canada

Flow-Based Generative Modeling MeanFlow One-Step Generative Modeling

Ask PDF

May 12, 2026

One-Step Generative Modeling via Wasserstein Gradient Flows

Diffusion models and flow-based methods have shown impressive generative capability, especially for images, but their sampling is expensive because it requires many iterative updates. We introduce W-Flow, a framework for training a generator that transforms samples from a simple reference distribution into samples from a target data distribution in a single step. This is achieved in two steps: we first define an evolution from the reference distribution to the target distribution through a Wasserstein gradient flow that minimizes an energy functional; second, we train a static neural generator to compress this evolution into one-step generation. We instantiate the energy functional with the Sinkhorn divergence, which yields an efficient optimal-transport-based update rule that captures global distributional discrepancy and improves coverage of the target distribution. We further prove that the finite-sample training dynamics converge to the continuous-time distributional dynamics under suitable assumptions. Empirically, W-Flow sets a new state of the art for one-step ImageNet 256

\times

256 generation, achieving 1.29 FID, with improved mode coverage and domain transfer. Compared to multi-step diffusion models with similar FID scores, our method yields approximately 100

\times

faster sampling. These results show that Wasserstein gradient flows provide a principled and effective foundation for fast and high-fidelity generative modeling.

Jiaqi Han, Puheng Li, Qiushan Guo +3

Stanford University · ByteDance

Wasserstein Gradient Flows Image Generation One-Step Generative Modeling Generative Modeling Flow-Based Generative Modeling

Ask PDF

Abstract

Diffusion Model Adaptation

Explore similar work

Continuous Adversarial Flow Models

Discrete Meanflow Training Curriculum

One-Step Generative Modeling via Wasserstein Gradient Flows