Tag
提出了Noise Policy Optimization（NPO），通过在初始噪声空间上学习一个摊销流策略，将少步生成模型与黑盒奖励对齐，无需微调生成器即可在图像、语言和蛋白质结构生成中实现更优的奖励-质量权衡，推理成本比测试时搜索低16至549倍。
提出 Routing Policy Optimization（RPO），在扩散生成器内部通过随机注意力路由引入探索，以对齐奖励信号，同时保留原始确定性采样器，并在文本到图像/视频生成上实现相比 Flow-GRPO 方法近 4 倍的训练效率提升。
GeoFix提出了一种统一框架，通过单一几何奖励提升4D生成中的几何一致性和视觉质量，该奖励衡量静态区域的跨视角深度一致性，并针对视角重叠效应进行校准，实现了WorldScore 3D一致性提升39%和VBench质量提升4%。
This paper introduces ZeNOVA, a gradient-free, zeroth-order initial noise optimization method for aligning deterministic one- or few-step generative models to black-box rewards, using annealed soft-value guidance, manifold-constrained hyperspherical Langevin dynamics, and Metropolis-Hastings jumping to achieve more stable optimization than existing approaches.
The paper introduces Reward-Weighted Transport Distillation (RWTD), a post-training method for one-step generative models that needs only generated samples and scalar rewards, blending tilted current and reference distributions via feature-space optimal transport. It boosts the GenEval score of the SanaSprint 1.6B one-step backbone from 0.73 to 0.80 while preserving compositional capabilities.
The paper introduces Wasserstein-Tilted Flow Maps (WTF), a simulation-free reinforcement learning algorithm for fine-tuning flow-based generative models to enhance reward alignment, achieving higher rewards with up to 280x less compute than baselines.
ReNFT is a method to repair mode collapse in reward post-training of diffusion generators by internally recalibrating probability mass, improving diversity while retaining high reward scores.
GRAS is a training-free method for reward alignment in discrete diffusion models that reduces variance in guided proposals and adaptively selects particles, achieving state-of-the-art results on DNA and protein design tasks.