Feed Entities Podcasts Deep Dives Developers About

[KongchangAI]

Real-time AI knowledge base. Every claim is traceable to its source, with data freshness under 24 hours.

Powered by 240+ first-hand AI sources

Knowledge

Feed
Hot Entities
Podcasts

Developers

CLI
MCP Server
REST API
About

© 2026 KongchangAI. The freshest, most accurate AI knowledge

Verified70% confidenceFactTime unknown

AI扩图每次生成的结果会有所不同，这种随机性源于扩散模型每次从不同的随机噪声出发

4

Sources

70%

Confidence

Long-term

Relevance

6/2/2026

First Seen

Sources

AI扩图功能详解：一键扩展图片边界

bilibili达内教育

Related Entities

Diffusion Modelconcept

Related Claims

Unverified扩散模型的迭代去噪过程对数值误差敏感，每步的微小误差会累积并影响最终图像，采样步数越多量化误差影响越大74% similar Unverified主体漂移本质上是模型在时序一致性建模上的不足，源于扩散模型每帧去噪引入随机性叠加时序注意力对长程依赖捕获能力有限74% similar Verified扩散模型通过逐步去噪从随机噪声中还原目标图像或视频帧，通常需要数十步去噪迭代，计算开销极大73% similar UnverifiedLLM的输出本质上具有随机性，相同的输入在不同调用中可能产生截然不同的结果71% similar Unverified扩散模型的推理是迭代去噪过程，通常需要数十甚至数百个时间步，对数值误差格外敏感71% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/36825

API

curl https://kongchang.com/api/v1/knowledge/claims/36825

MCP

get_claim(id=36825)