MeanFlow Transfer Accelerates Diffusion Model Adaptation.

Training fast generative models on new domains with limited data presents two main hurdles: adapting diverse pretrained diffusion models while maintaining acceleration, and extending adversarial refinement to MeanFlow (MF) models which predict average velocities over finite intervals. This research proposes a dual solution to these problems. First, MeanFlow-Transfer (MF-T) is introduced, which maps outputs from various source models (e.g., DiT, SiT, JiT) into a unified velocity representation. This allows for initializing an MF generator from source weights and optimizing an MF objective on the target domain, effectively combining adaptation and acceleration within a single training loop for a wide range of pretrained models. Second, Continuous Adversarial MeanFlow (CAMF) is presented as a post-training stage. CAMF extends continuous adversarial flow models to MF's finite-interval average velocities by contrasting changes in a learned potential between real and predicted interval endpoints. This approach recovers fine details that MF regression might average out. Experiments adapting four ImageNet-based models to five target domains show that MF-T with CAMF matches or surpasses fine-tuned teachers in quality (FID, FDD) with up to 125 times fewer Neural Function Evaluations (NFEs), and CAMF alone improves MF-T's few-step FID by an average of 29%.