视频｜Vidu3 参考生视频

The hardest problem in AI video isn't quality, it's consistency - the character in shot one looks nothing like the character in shot three. Vidu3's reference mode attacks that directly: you feed it up to seven reference images and address them by number in your prompt (@1 for the first, @2 for the second…), and the model keeps identity and appearance stable across the generated clip. Vidu is a real, active model line in the ecosystem (141 threads in the community corpus), and this node is Tikpan's wrapper over the /ent/v2/reference2video endpoint - the model ID is viduq3.

The README's framing is honest about the use case: this is for 角色一致性生视频 - keeping a character consistent so a brand mascot, an influencer's face, or a product hero survives from one video to the next. For e-commerce short-content pipelines, that's the difference between usable and unusable.

How it works

Build a payload with the prompt plus up to seven reference images (参考图1 is required and maps to @1; 2–7 optional), submit to the endpoint, then poll ent/v2/tasks/{task_id}/creations until the render finishes, and download the result. The prompt is the whole game here - the tooltip spells it out: use @1/@2/… in the prompt to reference the corresponding image, and the more concrete the scene/motion description, the better. English recommended.

Useful controls beyond the basics: 视频时长 (3–16s), 清晰度 (540p/720p/1080p), 画面比例 (16:9/9:16/1:1/4:3/3:4), 智能主体库 (the Vidu auto_subjects parameter - let the model auto-extract subject identities), 音画同步 (default on, generates matching ambient audio/music) with 音频类型 to constrain (All/Speech-only/Sound-effects-only), and 错峰生成 (off_peak - submit during off-peak for a discount, accepting a possibly delayed render).

Inputs and outputs that matter

A beginner's flow: API_密钥 → 参考图1 (required) → a prompt referencing it like the default @1 appears in a cinematic commercial video, natural motion, stable identity → pick duration/resolution/aspect → run. The local polling knobs 最长等待秒数 (default 1200) and 查询间隔秒数 (default 8) are just how long you'll wait; 校验HTTPS证书 and 跳过错误 are the standard local options. Outputs: 📁本地保存路径, 🆔任务ID, 🔗视频云端直链, 📄完整日志, 🎬视频输出 (VIDEO).

Installing

Pack install as usual - Manager search "Tikpan", or git clone https://github.com/htrert/ComfyUI-Tikpan-Pro into custom_nodes, restart. Key from tikpan.com.

Common issues

• "至少需要 1 张参考图". Reference mode needs at least one image - the node returns this error (or an empty result with 跳过错误 on) if 参考图1 is missing. This is the #1 mistake.
• Prompt not referencing any @N. If your prompt never mentions @1, the reference image has nothing to anchor to and you've basically paid for a plain text2video with extra upload cost. Reference your images by number.
• Too many references fight each other. Seven reference images with vague prompts produce mush. The character-consistency principle: fewer, high-quality, consistent references beat a pile of mixed ones (character-consistency.md).
• Long waits on big renders. 16s at 1080p is a real cloud job; if it times out at your 最长等待秒数, note the task ID and collect it with the Task Fetcher node.
• Seed hidden, no determinism. Same as the pack's other video nodes - the seed isn't exposed because it isn't a reliable reproduction knob through this relay. Don't chase exact re-runs; change seeds by changing prompt phrasing.