Continuous Adversarial MeanFlow Transfer

Yara Bahram* 1,2,3, Zahra Dehghani* 1,2,3,4, Mélodie Desbos 1,2,3
Eric Granger 1,2,3, Pablo Piantanida 2,4,5, Mohammadhadi Shateri 1,2,3
1LIVIA, 2ILLS, 3ÉTS Montréal, 4Mila – Quebec AI Institute
5CNRS, CentraleSupélec - Université Paris-Saclay
* Equal contribution
📚 arXiv 💻 Code
TL;DR: We transfer heterogeneous pretrained diffusion and flow models to new target domains as high-quality few-step MeanFlow generators, and further improve generation quality with continuous adversarial refinement.
\(x\ /\ \epsilon\ /\ v\ /\ u\)
Supports heterogeneous source parameterizations
Up to 125×
Fewer NFE with comparable or better FID and FDD than teacher fine-tuning
29%
Average improvement in Few-step FID in the CAMF stage
Overview of MeanFlow-Transfer and CAMF

MeanFlow-Transfer (MF-T) transfers heterogeneous pretrained diffusion and flow models to few-step MeanFlow generators, while Continuous Adversarial MeanFlow (CAMF) further improves their few-step generation quality through adversarial refinement. Together, MF-T + CAMF adapts a source-domain diffusion or flow model with any output parameterization into a high-quality few-step MeanFlow generator in a target domain, with \(\mathrm{NFE}_{\mathrm{trg}} \ll \mathrm{NFE}_{\mathrm{src}}\) (\(\mathrm{NFE}_{\mathrm{trg}} \sim \mathrm{NFE}_{\mathrm{src}}\) when the source model is already a MeanFlow model).

Target-Domain Generations

Dataset
iMF-XL/2
\(u\)-prediction
\(\mathrm{NFE}_{\mathrm{src}\rightarrow\mathrm{trg}} = 2 \rightarrow 2\)
SiT-XL/2
\(v\)-prediction
\(\mathrm{NFE}_{\mathrm{src}\rightarrow\mathrm{trg}} = 250 \times 2 \rightarrow 4\)
DiT-XL/2
\(\epsilon\)-prediction
\(\mathrm{NFE}_{\mathrm{src}\rightarrow\mathrm{trg}} = 250 \times 2 \rightarrow 8\)
JiT-H/16
\(x\)-prediction
\(\mathrm{NFE}_{\mathrm{src}\rightarrow\mathrm{trg}} = 50 \times 2 \rightarrow 4\)
CUB
CUB iMF CUB SiT CUB DiT CUB JiT
Cars
Cars iMF Cars SiT Cars DiT Cars JiT
ArtBench
ArtBench iMF ArtBench SiT ArtBench DiT ArtBench JiT

Target-domain generated images using MF-T + CAMF initialized from different source models. Images are uncurated.

Method

Our framework consists of two stages. MeanFlow-Transfer (MF-T) provides a common interface for pretrained diffusion and flow models with heterogeneous source parameterizations. Source models predicting \(x\), \(\epsilon\), \(v\), or \(u\) are mapped into a shared instantaneous-velocity representation, which is used to initialize and adapt a few-step MeanFlow generator \(u_\theta\) to the target domain. Continuous Adversarial MeanFlow (CAMF) then refines the adapted MF model by extending continuous adversarial learning from instantaneous velocities to the finite-interval average velocity predicted by MeanFlow.

MeanFlow-Transfer (MF-T)

MeanFlow-Transfer

MeanFlow-Transfer (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 \(u_\theta\).

Continuous Adversarial MeanFlow (CAMF)

Continuous Adversarial MeanFlow

Continuous Adversarial MeanFlow (CAMF). The MF model transports \(z_t\) to the predicted endpoint \(\hat{z}_r = z_t - (t-r)u_\theta\), 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 \(\mathcal{A}_{\mathrm{fake}}\).

Citation

If you find this work useful, please consider citing: 

@misc{bahram2026continuousadversarialmeanflowtransfer,
  title={Continuous Adversarial MeanFlow Transfer},
  author={Yara Bahram and Zahra Dehghani and Mélodie Desbos and Eric Granger and Pablo Piantanida and Mohammadhadi Shateri},
  year={2026},
  eprint={2608.19540},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2608.19540},
}