[论文] InstantFusion: A Shared Latent Interface for Heterogeneous Image Gener...

论文概要

研究领域: CV
作者: Xuanlang Dai, Jiazi Bu, Yujie Zhou, Yuhong Liu, Beichen Zhang, Penghui Yang, Ziyu Liu, Yuhang Zang
发布时间: 2026-10-09
arXiv: 2610.13944

中文摘要

扩散/流模型已成为图像生成的事实标准范式，催生了架构与表征日益异构的现代生成器家族。尽管生成能力卓越，但跨模型的潜空间对齐与去噪动力学一致性仍缺乏探索。本工作揭示了一个有趣的跨模型兼容性：独立训练的生成器展现出相似的潜空间空间组织，且在未对齐模型之间直接交接中间状态仍能保留粗略视觉内容。基于此提出 InstantFusion——轻量级共享潜空间接口，连接异构图像生成器的去噪轨迹。具体地，引入模型特定的潜空间自编码器（LAE），将不同生成器的中间状态投影到共享潜空间，再解码到目标生成器的潜空间；通过两个互补目标约束：对齐来自同一图像在匹配噪声级别的中间状态、同步其对应的局部去噪速度。训练完成后，去噪可在冻结生成器间无缝切换，无需为每对模型训练专门映射。InstantFusion 在加速、多偏好组合、跨架构在策略蒸馏和跨模型组合等场景中均展现竞争力。

原文摘要

Diffusion/flow models have emerged as the de facto paradigms for image generation, giving rise to a diverse family of modern generators with increasingly heterogeneous architectures and representations. Despite their remarkable generation capabilities, the cross-model alignment of latent representations and the consistency of their underlying denoising dynamics remain underexplored. In this work, we uncover an intriguing cross-model compatibility: independently trained generators exhibit similar latent spatial organization, while directly handing off intermediate states between unaligned models can still preserve coarse visual content. Motivated by these observations, we propose InstantFusion, a lightweight shared latent interface that connects the denoising trajectories of heterogeneous i...

自动采集于 2026-10-13

#论文 #arXiv #CV #小凯