Unverified50% confidenceTradeoffExact time
在多角色复杂场景中，不同角色的特征可能相互干扰（Feature Leakage），导致一个角色的面部特征泄漏到另一个角色上
1
Sources
50%
Confidence
Long-term
Relevance
8/10/2026
First Seen
Sources
GPT-2+Seedance 2.5实测：AI黑暗奇幻战斗片能力边界在哪
Related Claims
Unverified区域提示技术在多个角色LoRA同时加载时，因不同LoRA的低秩矩阵方向冲突同时修改同一批注意力层权重，导致角色面部融合或特征错位64% similarUnverified2D游戏中碰撞体与视觉体分离是常见做法，解决角色图片含透明区域导致的碰撞判定不准问题，方案是缩小碰撞框只让下半部分参与判定61% similarUnverifiedCharacter consistency is a widely recognized technical challenge in AI visual creation, as mainstream diffusion models introduce randomness in each generation that causes drift in facial features, body shape, and skin tone.59% 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58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/722317API
curl https://kongchang.com/api/v1/knowledge/claims/722317MCP
get_claim(id=722317)