Unverified50% confidenceFactExact time
参考视频在技术层面充当「结构先验」角色，纹理细节和光照风格由文本提示词和模型自身预训练知识填充
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
Related Claims
UnverifiedRepresentative academic works in reference-driven video generation include Animate Anyone, Champ, and MimicMotion, which achieve the separation of identity and motion by introducing dedicated reference encoders (ReferenceNet) or appearance attention mechanisms.79% similarUnverifiedReference video templates extract motion information (such as optical flow, pose skeletons, and depth information) as conditional constraints for the generation process; common implementation methods include ControlNet, IP-Adapter, and Motion LoRA.73% similarUnverifiedIC-LoRA 借鉴大语言模型的 In-Context Learning 范式，将参考图像或视频帧作为条件信号，通过跨注意力机制影响目标帧的生成72% similarUnverifiedThe experiment uses a DiffusionGemma custom node paired with the LTX 2.3 video generation model, with inputs consisting of one static reference image, one reference video, and one natural language prompt.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/490437API
curl https://kongchang.com/api/v1/knowledge/claims/490437MCP
get_claim(id=490437)