Concurrent Submit | Zhenzhen_vidu_ref2video

This is the Vidu node for when "a character" isn't enough - you have characters, plural, and they need to hold identity while a scene plays out. Vidu's reference-to-video mode takes up to seven reference images and lets you tag up to three of them as subjects, each with its own voice. That's a real multi-character pipeline: subject 1 gets reference image 2 and a male voice, subject 2 gets reference image 5 and a female voice, and the prompt choreographs them together. It's the closest thing in this pack to a two-hander script writer's tool, and the concurrent wrapper means you can run several multi-subject takes in parallel.

Inputs that matter

• prompt - required. Describe the action, referencing the subjects.
• model - viduq2 (default), viduq1, vidu2.0, vidu1.5.
• image1–image7 - up to seven reference images. The subjectN_id fields tell Vidu which of these are the actors: subject1_id/subject2_id/subject3_id default to "1", "2", "3" (i.e. image slots 1–3).
• subjectN_voice_id - per-subject voice, chosen from the same ~300-voice catalog as the img2video node. Leave empty and the subject is silent.
• audio - boolean for generating audio on the clip.
• duration - 1 to 10 seconds, default 5.
• aspect_ratio - 16:9, 9:16, 4:3, 3:4, or 1:1.
• resolution - 540p, 720p (default), 1080p.
• movement_amplitude, bgm, off_peak, watermark, wm_position - same controls as the rest of the Vidu family.
• api_key, seed, skip_error - standard.

Output: one COMFLY_VIDEO_FUTURE task.

How it works

The wrapper is the standard ComflyConcurrent.py pattern: validate, push the real Comfly_vidu_ref2video call onto the bounded video executor (ten workers), return the future immediately. The Concurrent Collect Videos node waits on all tasks, restores slot order, and reports per-slot status; failure_mode: placeholder converts a failed slot into a blank clip rather than aborting the run.

Setup

One-time install: ComfyUI Manager (search "Comfyui-zhenzhen"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

No weights. You need a Zhenzhen key from ai.t8star.org (overseas) or api.seedance.nz (domestic). Reference-to-video with multiple subjects and voices bills higher than a plain text-to-video clip - more inputs, more processing - so check the per-call cost before you queue a wall of takes.

Gotchas

Character consistency is the entire risk surface. Vidu's subject anchoring is good but not actor-perfect: faces drift across a long clip, and the more subjects you tag, the more the model has to juggle. Keep subject count low (two reads better than three), give each subject one clean, same-perspective reference image, and don't expect the same character to survive being referenced from wildly different angles. Also note subjectN_id is a string, not a dropdown - "1" means image slot 1, and a stale ID silently references the wrong image. And as everywhere in this pack, timeouts aren't guaranteed failures: check your async task page.