跳到主要内容Skip to main content

参考生视频Reference-to-Video

Reference-to-Video

  • 图像语音视频Image, speech & video
  • 第三段 · 用深Stage 3 · Going Deeper
  • 引用核验 2026-11-19refs checked 2026-11-19
解决什么问题What problem it solves

用几张图、一段视频或一段音频当参考,生成主体、风格、镜头一致的新视频。Use several images, a video or an audio clip as references to generate a new video with consistent subject, style and camera.

这是什么What it is

模型不只读文字提示,还读参考素材里的身份、服装、构图、运镜与音色。

与图生视频的区别:图生视频通常只给首帧;参考生视频给的是「要保留什么」的约束。

参考给得越多,控制越强,但冲突也越多;素材之间不一致时,模型会自行取舍。

验收要看参考保持度与可编辑性,不只看画面漂亮。

The model reads not only a text prompt but identity, clothing, composition, camera movement and timbre from reference material.

Unlike image-to-video, which usually supplies only a first frame, reference-to-video supplies constraints on what must be preserved.

More references give stronger control but also more conflicts; when references disagree, the model makes its own trade-offs.

Acceptance should check reference fidelity and editability, not only whether the output looks good.

为什么需要知道Why it matters

Seedance 2.5、Veo 等都在往多模态参考走;这是「生成一段」到「完成一支片子」的分水岭。Seedance 2.5, Veo and others are moving toward multimodal references; this is the step from generating a clip to finishing a piece.

出处Sources
  • 字节 Seed · Seedance 2.5ByteDance Seed · Seedance 2.5 ↗
  • Google DeepMind · VeoGoogle DeepMind · Veo ↗
  • 学 AI · 视频生成Learn AI · Video generation ↗