Unverified75% confidenceOpinionExact time
扩散模型本质上是对像素分布的学习，缺乏对因果物理过程的建模
1
Sources
75%
Confidence
Long-term
Relevance
5/29/2026
First Seen
Sources
Gemini Omni: How Powerful Is AI Video Generation That Understands Physics?
twitterGeminiApp5/29/2026
Related Entities
Related Claims
Unverified扩散模型的归纳偏置天然倾向于平滑、渐变的输出，无法在架构层面强制施加网格约束，即便专门训练的LoRA模型也只能在统计意义上让输出更像像素画77% similarUnverified当前扩散模型对非标品或定制化产品往往进行创意性补全而非精准还原，因为模型学习的是视觉分布而非具体商品的精确外观77% similarUnverified扩散模型在连续像素空间中学习概率分布，而非在离散符号空间中建立语义映射，与GPT系列语言模型的Token化处理路径存在架构分歧73% similarUnverifiedTraditional diffusion models frequently make errors when generating text, with the root cause being that the training objective is to reconstruct the pixel distribution of images from noise, rather than to understand the semantic structure of characters.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/184API
curl https://kongchang.com/api/v1/knowledge/claims/184MCP
get_claim(id=184)