Improved Distributional Diffusion Models

• 类型：arxiv
• 标识：2609.37147
• 链接：https://arxiv.org/abs/2609.37147
• 主分类：multimodal
• 形态：method
• TLDR：Distributional Diffusion Models (DDMs) replace the standard mean-prediction denoiser with a distributional denoiser trained via a scoring rule objective, learning a stochastic approximation to p(x_1 mid x_t) rather than its conditional mean. However, scaling DDMs to modern image-generation settings faces two obstacles: (i) multi-particle training incurs overhead that scales with the number of particles, (ii) DDMs use globally fixed scoring rule hyperparameters, forcing a single trade-off across sampling budgets. We mitigate these limitations by deferring particle expansion to late transformer
• 待LLM分类：否
• 标题中文：[标题中文] 改进的分布式扩散模型
• TLDR中文：[TLDR中文] 分布式扩散模型（DDM）用通过 scoring rule 目标训练的分布式去噪器替代标准的均值预测去噪器，学习对 p(x_1 mid x_t) 的随机近似而非其条件均值。然而，将 DDM 扩展到现代图像生成场景面临两大障碍：(i) 多粒子训练带来随粒子数增长的额外开销；(ii) DDM 使用全局固定的 scoring rule 超参数，迫使在所有采样预算下采用单一权衡。我们通过将粒子扩展推迟到 Transformer 后期
• 来源文件：
• /inbox/tom/_candidates/2026-09-30-agent-rag-longcontext-candidates.json