few-step-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。

本文提出 data-space iteration 框架，将少步生成的迭代过程从概率流空间移至数据空间，从而完全移除对流离散化时间表的依赖，并在 ImageNet 256x256 类条件生成上以最小改动集成 DMD，超越了需要手动调参时间表的基线方法。

论文提出 E-MoE（Enhanced Mixture-of-Experts），用 MoE 路由决策构成的离散共享隐变量来建模非因子化的反向过程，缓解连续隐变量的后验坍缩问题，并在无需增加活跃参数的情况下提升掩码扩散语言模型在少步生成场景下的样本质量。

This paper introduces untied self-conditioning to correct train-inference mismatch in flow-matching language models, improving few-step generation quality with lower perplexity and no retraining.

This paper proposes MeanFlow-Transfer (MF-T) and Continuous Adversarial MeanFlow (CAMF) to unify the adaptation and acceleration of pretrained diffusion and flow models, enabling high-quality few-step generation on new domains with limited data.

The paper introduces Latent-Kernel Discrete Flow Maps (LKF), a flow-map kernel for discrete diffusion models that captures correlations between positions via a shared latent, enabling few-step generation without distillation and improving text generation perplexity.

Perceptual Flow Matching supervises flow matching in perceptual feature space, enabling high-quality few-step generation with 4-8 sampling steps instead of 35-50, without needing teacher models.

The paper identifies why deterministic few-step generation fails for text while succeeding for images: the sharp categorical readout in text decoders amplifies small errors, causing token flips, whereas continuous image decoders are smooth. It proposes diagnostics (DABI, CCI) and escape mechanisms such as categorical commitment and stochastic re-injection.

VESFlow is a training-free safety method for flow matching-based text-to-image generation that edits velocity fields to ensure safe output while maintaining prompt integrity.

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.

FlowLM introduces a flow matching language model derived from pre-trained diffusion models via efficient fine-tuning, enabling high-quality few-step text generation that rivals 2,000-step diffusion sampling with far fewer training epochs.