distribution-matching

FlashVSR+ 提出基于混合注意力的高效视频超分辨率架构与 Reward-Yield 分布匹配蒸馏（Ry-DMD），实现 2 倍 DiT 加速、45% 峰值显存降低，可在单张 24GB RTX 4090 上完成 4K 视频超分，并支持黑盒奖励对齐的单步生成。

本文证明了 Drifting Models 与 Distribution Matching Distillation (DMD/DMD2) 在优化目标上的等价性，并基于此视角提出了改进的 Multi-Bandwidth Distribution Matching Distillation (MBDMD) 方法，用于提升一步生成模型的性能。

论文提出 Consistent Distribution Matching (CDMD)，通过在单一学生网络内统一样本生成与速度估计，实现无需辅助 fake-score 网络的数据自由扩散蒸馏，在 ImageNet 256×256 上以 1 NFE 达到 2.04 FID，优于 DMD2 等基线，并给出基于 Wasserstein 收敛的理论分析。

Salt++ 是一个两阶段后训练框架，通过 CausalSelf-Flow 和上下文对齐的自回归 DMD，解决少步流式音视频生成中教师强制与因果分布蒸馏的上下文错配问题，在 480p 4-step 设定下相对 OmniForcing 提升视觉质量 57%、运动质量 45%。

RTDMD is a two-stage framework combining distribution matching distillation with reward-guided reinforcement learning to improve few-step image generation alignment with human preferences. It achieves state-of-the-art results on multiple models with only 4 inference steps.

This paper introduces Domain Generalizable Dataset Distillation (DGDD), a new problem setting that targets out-of-distribution generalization of distilled datasets, and proposes Spectral Gradient Surgery (SGS) to disentangle class-discriminative and domain-specific information by leveraging cross-domain gradient agreement in the spectral domain.

This paper identifies mode collapse in on-policy RL methods like GRPO and proposes DMPO, which approximates forward KL minimization to maintain solution diversity. It achieves significant improvements on NP-hard combinatorial optimization and mathematical reasoning tasks.