Zhenzhen_vidu_ref2video

Vidu's reference-to-video mode is the one where you hand it a character still and it keeps that character consistent while the scene moves - no ControlNet, no IPAdapter, no local weights at all, because Comfly_vidu_ref2video is a paid API client for the Vidu relay. That's the whole appeal: character consistency without fighting a local identity stack. You feed up to seven reference images, optionally tell it which subject each image belongs to, and even assign the subject a voice so the generated character can talk. This is about as close to "generate a scene with my OC" as a ComfyUI node gets without local training.

Inputs that matter

• prompt - the scene description. Required.
• model - viduq2 (default), viduq1, vidu2.0, or vidu1.5. Newer isn't always better for your exact job; q2 is the current default for a reason.
• image1…image7 - your reference stills. Fewer is fine; the subject consistency comes from Vidu, not from quantity.
• audio - toggles whether the output gets a generated voice track.
• subject1_id … subject3_id - map each reference image to a subject slot (defaults "1", "2", "3").
• subjectN_voice_id - the kicker: hundreds of voices, from female-shaonv to the male-qn-* premium set. Assign one per subject and your character talks.
• duration - 1–10 seconds.
• aspect_ratio / resolution - 16:9/9:16/4:3/3:4/1:1 and 540p/720p/1080p.
• movement_amplitude - auto/small/medium/large; large is where the shot gets lively.
• bgm - background music track on/off.
• off_peak - run during off-peak hours for a cheaper rate (Vidu's discount window).
• watermark / wm_position - if you don't want Vidu's watermark baked in, keep it off.

Outputs: video (into ComfyUI's native Save Video), plus video_url, task_id, response.

Install

Standard Comfly:

cd ComfyUI/custom_nodes
git clone https://github.com/ainewsto/Comfyui_Comfly

Or ComfyUI Manager → "Comfyui_Comfly" → restart. Nothing runs locally; you need credits at https://ai.comfly.chat and the key in Comflyapi.json (or the node's api_key override).

Gotchas

Voice assignment is the fiddly part - the subject ids are strings ("1", "2", "3") and if your images aren't clearly separable subjects, the voice lands on the wrong one and you've paid for a weird result. And remember the relay is a third-party service: a render that fails still cost you a slot, so keep off_peak in mind for batch work and use the task_id/response outputs to check what actually happened before re-queuing. The watermark toggle defaults off, which is nice - just don't assume it stays off if you load an example workflow someone else saved.