论文证明Drifting模型与分布匹配蒸馏等价并提出MBDMD
Get the story
2026年10月9日，arXiv Machine Learning Theory 发布研究《Multi-Bandwidth Distribution Matching Distillation》。该研究证明，把预训练 Diffusion 与 Flow Style 生成模型（DFSGMs）的速度场/噪声场转换为 Drifting Models 的吸引力场，并从生成分布估计排斥力场后，训练 Drifting Model 与 Distribution Matching Distillation（DMD/DMD2）天然等价。这是目前该事件唯一报道，尚未见后续验证或争议信息。
Generated from reports · updated 3 hr ago
Timeline
Follow the coverage from different angles.
- arXiv · Machine Learning TheoryMulti-Bandwidth Distribution Matching Distillation：证明 Distribution Matching Distillation 与 Drifting Models 等价
研究证明，把预训练 Diffusion 与 Flow Style 生成模型（DFSGMs）的速度场/噪声场转换为 Drifting Models 的吸引力场，并从生成分布估计排斥力场后，训练 Drifting Model 与 Distribution Matching Distillation（DMD/DMD2）天然等价。
Heat trend
Current heat 9·Comparable peak 10（Oct 9）·Comparable change over 24 hours –
The trend compares only the same participants observed continuously; its range may be smaller than the current heat count. Move or click on the chart to inspect hourly heat; use the left and right arrow keys to switch.