Unverified50% confidenceFactExact time
原生视觉能力指模型在预训练阶段就将图像、视频等多模态数据与文本数据联合训练，而非通过后期插件式适配器拼接视觉编码器
1
Sources
50%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
Unverified多模态实现路径分为早融合（预训练阶段混合多模态数据）和晚融合（先用独立视觉编码器ViT提取图像特征，再通过投影层对齐到语言模型embedding空间）71% similarUnverified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70% similarUnverifiedDINOv2 is a visual foundation model released by Meta AI in 2023, pre-trained on large-scale unlabeled image data using a self-supervised learning paradigm.70% 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68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/514804API
curl https://kongchang.com/api/v1/knowledge/claims/514804MCP
get_claim(id=514804)