Unverified50% confidenceTradeoffExact time
ControlNet通过引入与主干U-Net并联的可训练副本，将姿态骨架图、深度图等结构化控制信号注入生成过程，但对面部特征的精确复现能力有限
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
Unverified原生多模态预训练在统一token空间中对图像patch和文本token进行联合自回归训练，使模型在权重层面形成跨模态语义绑定65% similarUnverified预训练的U-Net通常指编码器部分使用了在ImageNet等大规模数据集上预训练的骨干网络（如ResNet、EfficientNet），仅解码器部分从零训练65% similarUnverifiedPerson re-identification models such as OSNet and FastReID are mostly trained on surveillance pedestrian datasets, and applying them directly to film scenes will result in domain shift issues.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/536229API
curl https://kongchang.com/api/v1/knowledge/claims/536229MCP
get_claim(id=536229)