Unverified50% confidenceFactExact time
DiffusionGemma是首个在大规模（26B参数）上证明扩散范式可以匹敌自回归质量的工作
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/27/2026
First Seen
Valid until: 11/25/2026
Sources
Related Entities
Related Claims
UnverifiedDiffusionGemma复用现成的Gemma系列（约26B）模型权重，通过监督微调和采样器蒸馏与强化学习两步改造完成转换80% similarUnverified扩散模型的数学本质是基于随机微分方程和得分匹配理论，支撑着Stable Diffusion、DALL-E等图像生成系统74% similarUnverifiedDALL·E 2和Stable Diffusion均基于扩散模型原理71% similarVerifiedStable Diffusion is built on the Latent Diffusion Model (LDM) architecture70% similarUnverified扩散模型生成过程在浮点数连续空间中运算，只在最终步骤才量化为整数，因此归纳偏置天然倾向于平滑渐变输出而非硬边界69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/807338API
curl https://kongchang.com/api/v1/knowledge/claims/807338MCP
get_claim(id=807338)