Unverified50% confidenceFactExact time
人脸替换模型通常以正面或近正面人脸为主要训练数据，对大角度侧脸和局部遮挡的生成质量显著下降
1
Sources
50%
Confidence
Long-term
Relevance
7/18/2026
First Seen
Sources
Related Claims
UnverifiedCurrent mainstream AI face-swapping models are prone to image stretching, twitching, or misalignment when encountering occlusion or large-angle profile views.80% similarUnverified现代人脸识别系统通常分为人脸检测和人脸识别/比对两个独立阶段72% similarUnverified作者设计损失函数直接计算生成图像的人脸嵌入与目标人脸嵌入之间的距离，实现端到端可微的人脸相似度优化69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/552369API
curl https://kongchang.com/api/v1/knowledge/claims/552369MCP
get_claim(id=552369)