multi-modal-generation

提出简化匹配蒸馏（SMD）框架，通过纯回归的on-policy速度回归与Pivot残差代理，实现无需辅助评分模型或对抗训练的4步少步扩散生成，在图像、3D、视频及音视频联合生成任务上超越强基线并可扩展至33B参数。

MMDiff extends frozen diffusion transformers into multi-modal generative systems using lightweight decoders, achieving significant improvements in semantic segmentation and other perceptual tasks through multi-timestep feature fusion.