Research Papers 论文研究 1d ago Updated 20h ago 更新于 20小时前 47

Continuous Adversarial MeanFlow Transfer 连续对抗平均流迁移

MeanFlow-Transfer (MF-T) unifies domain adaptation and sampling acceleration by mapping heterogeneous pretrained model outputs (DiT/ε, SiT/v, JiT/x, iMF/u) into a shared velocity representation for initialization Continuous Adversarial MeanFlow (CAMF) extends adversarial refinement from instantaneous velocities to MeanFlow's finite-interval average velocities, recovering fine details that regression averaging loses MF-T with CAMF matches or exceeds fine-tuned teacher models in FID and FDD at up 提出MeanFlow-Transfer (MF-T)框架,将异构预训练扩散/流模型(DiT/ε, SiT/v, JiT/x, iMF/u)统一映射到共享速度表示,实现适配与加速的联合优化 引入Continuous Adversarial MeanFlow (CAMF)后训练阶段,将对抗精炼从瞬时速度扩展到MeanFlow的有限区间平均速度,恢复被回归平滑掉的细节 在四个ImageNet源模型适配到五个目标域的实验中,MF-T+CAMF以最多125倍更少的NFEs匹配或超越微调教师模型的FID和FDD指标 CAMF平均将MF-T的少步FID提升29%,在区间趋近零时退化为瞬时对抗准则

62
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • MeanFlow-Transfer (MF-T) unifies domain adaptation and sampling acceleration by mapping heterogeneous pretrained model outputs (DiT/ε, SiT/v, JiT/x, iMF/u) into a shared velocity representation for initialization
  • Continuous Adversarial MeanFlow (CAMF) extends adversarial refinement from instantaneous velocities to MeanFlow's finite-interval average velocities, recovering fine details that regression averaging loses
  • MF-T with CAMF matches or exceeds fine-tuned teacher models in FID and FDD at up to 125× fewer Neural Function Evaluations across five target domains
  • CAMF improves MF-T's few-step FID by 29% on average, demonstrating significant quality gains at drastically reduced computational cost
  • The framework addresses the lack of a common acceleration target across heterogeneous pretrained diffusion/flow parameterizations

Why It Matters

This work directly tackles two persistent bottlenecks in deploying generative models: the high sampling cost of pretrained diffusion/flow models and the difficulty of adapting them to new domains with limited data. By unifying adaptation and acceleration into a single training loop that works across heterogeneous model parameterizations, it removes a major barrier for practitioners who need to deploy fast, high-quality generators in resource-constrained settings.

Technical Details

  • MeanFlow-Transfer (MF-T): Maps outputs from four different pretrained source models—DiT (ε-parameterization), SiT (v-parameterization), JiT (x-parameterization), and iMF (u-parameterization)—into a shared velocity representation, then initializes a MeanFlow generator from source weights and optimizes the MF objective on the target domain
  • Continuous Adversarial MeanFlow (CAMF): A post-training stage that formulates adversarial refinement over finite-interval average velocities rather than instantaneous velocities; it contrasts changes in a learned potential between real and predicted interval endpoints, recovering high-frequency detail that MF regression averages away
  • Theoretical consistency: CAMF reduces to the instantaneous adversarial criterion in the vanishing-interval limit, ensuring theoretical coherence with prior work
  • Evaluation: Adapted four ImageNet-based source models to five target domains, measuring performance via FID and FDD metrics at varying Neural Function Evaluation (NFE) budgets

Industry Insight

  • The unified velocity representation approach could become a standard bridge for transferring capabilities across the increasingly heterogeneous landscape of diffusion and flow-based generative models, reducing the need for model-specific adaptation pipelines
  • The 125× NFE reduction with competitive or superior quality opens practical pathways for deploying high-fidelity generators on edge devices or in real-time applications where computational budget is severely constrained
  • The extension of adversarial refinement to finite-interval velocities provides a generalizable framework that could be adapted to other flow-based generative modeling paradigms beyond MeanFlow

TL;DR

  • 提出MeanFlow-Transfer (MF-T)框架,将异构预训练扩散/流模型(DiT/ε, SiT/v, JiT/x, iMF/u)统一映射到共享速度表示,实现适配与加速的联合优化
  • 引入Continuous Adversarial MeanFlow (CAMF)后训练阶段,将对抗精炼从瞬时速度扩展到MeanFlow的有限区间平均速度,恢复被回归平滑掉的细节
  • 在四个ImageNet源模型适配到五个目标域的实验中,MF-T+CAMF以最多125倍更少的NFEs匹配或超越微调教师模型的FID和FDD指标
  • CAMF平均将MF-T的少步FID提升29%,在区间趋近零时退化为瞬时对抗准则

为什么值得看

本文解决了扩散模型领域长期存在的两个关键痛点:异构预训练模型的统一适配困难,以及少步采样质量下降问题。提出的框架为不同参数化的生成模型提供了通用的加速路径,对抗精炼的扩展也为少步生成质量提升提供了新思路。

技术解析

  • MF-T框架:将不同参数化(ε/x/v/u)的源模型输出映射到共享速度表示,用源权重初始化MeanFlow生成器,在目标域上优化MF目标函数,统一了适配与加速过程
  • CAMF方法:通过对比真实样本与预测样本在区间端点的学习势函数变化来构建对抗损失,恢复MeanFlow回归平均掉的细粒度细节,理论上在vanishing-interval极限下退化为瞬时对抗准则
  • 实验设置:四个ImageNet预训练源模型(DiT, SiT, JiT, iMF)适配到五个目标域,使用FID和FDD作为评估指标,NFEs减少达125倍
  • 性能表现:MF-T+CAMF在少步采样下匹配或超越微调教师模型,CAMF平均提升29%的少步FID

行业启示

  • 为预训练扩散模型的领域适配提供了统一解决方案,降低了不同参数化模型间的迁移成本,有利于构建可复用的生成模型基础设施
  • 少步采样质量的显著提升意味着实际部署时可在保持生成质量的同时大幅降低计算开销,对边缘设备部署和实时应用具有重要价值
  • 对抗精炼与MeanFlow的结合为未来研究开辟了新方向,特别是在平衡生成速度与细节保真度方面提供了可借鉴的范式

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Image Generation 图像生成 Training 训练 Inference 推理 Research 科学研究