comfyui_vidu_api/参考生视频

Here's the node for the problem that eats video people alive: keeping one character consistent across shots. ViduReference2Video takes one to seven reference images of your subject, and Vidu's job is to make that subject - not a random doppelgänger - appear in the generated clip. It's the pack's character-consistency workhorse, and if you're making a story with a recurring character, this is the one you'll actually live in.

The multi-reference setup is the whole point. A single reference can be ambiguous - is it the outfit that matters, the face, the hairstyle? Throw three to seven shots of the same subject at the model (front, side, action shot) and it has a much better chance of locking the identity, which is the same reasoning that drives the whole identity-preservation side of this ecosystem: the more grounded references you give a model, the less it has to invent. That said, this is still a closed API - Vidu runs the model on its servers, so "your character" is defined by what the API accepts, not by a LoRA you trained.

How it works. Same pack machinery as the rest: key from api.json, upload every provided reference image through the presigned upload flow, submit /ent/v2/reference2video with the image list, poll every 5 seconds, download the clip, return it as VIDEO. The node only uploads the sockets you actually wired - 参考图_1 is required, 参考图_2 through 参考图_7 are optional, and it refuses to run with zero images.

Inputs that matter. Beyond the images themselves:

• 运行配置 - viduq1 - 5秒, vidu2.0 - 4秒, vidu1.5 - 4秒, vidu1.5 - 8秒. Note the author's own warning buried in the source: viduq1 is listed here but was being rejected by this endpoint in their testing, so if you hit an error on it, don't fight it - switch profiles.
• 提示词 - what the subject does ("a cute character running" is the default). Keep it short; the references carry the identity, the prompt carries the action.
• 宽高比 - this one has it (16:9 / 9:16 / 1:1), unlike ViduImage2Video.

Outputs: video, 封面链接 (cover URL), 任务ID - the standard trio, and the same failure behavior (error text on the string outputs instead of an exception; check the [Vidu::ViduReference2Video] console logs).

Install is the pack-standard, and there's nothing heavy in it:

cd ComfyUI/custom_nodes
git clone https://github.com/josephxie1/comfyui_vidu_api_node
# restart, then create api.json beside the nodes:
#   { "api_key": "vda_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" }

Key from platform.vidu.cn, or ComfyUI Manager → search comfyui_vidu_api_node. Only dependency is requests - no models, no VRAM.

Where people get burned. Two resolution traps are pre-validated for you, and they're easy to hit: viduq1 - 5秒 only accepts 1080p, and vidu2.0 - 4秒 caps at 720p (the reference endpoint won't do 1080p on that profile). The node refuses mismatches before spending your credits, which is a favor - though a confusing one the first time. Otherwise the pack-wide warnings all apply: no api.json means every node dies at startup, you need a ComfyUI build with comfy_api for the VIDEO output to exist, and the model line is frozen at Vidu Q1/1.5/2.0 while Vidu's current generation has moved on. And because this node sends up to seven of your images plus your prompt to a third-party server, it's the privacy-worst node in the pack - if those reference images are of a real person, remember where they're going.