comfyui_vidu_api/首尾帧生视频

This is the interpolation node: you provide the first frame and the last frame, and Vidu invents a plausible, smooth path between them. ViduStartEnd2Video is the pack's answer for the shots you can't nail with a prompt - the camera push from a wide establishing shot to a close-up, the object that needs to land exactly where you need it, the transition where the endpoint matters more than the journey. If the destination frame is non-negotiable, this is the node.

It sits in a different mental bucket from the rest of the pack. ViduText2Video and ViduImage2Video hand the model almost total creative control; this one gives it two anchor points and asks it to be a competent in-betweener. That's precisely the workflow motion designers reach for when they want a shot to end on a specific composition - you generate or grab both keyframes, wire them in, and Vidu does the 4–8 seconds of transition you'd otherwise be fighting with a text prompt to describe.

How it works. Standard pack machinery with one extra upload: both images go through the presigned upload flow, then a job is submitted to /ent/v2/start-end2video with the two image URIs, polled every 5 seconds, and the finished mp4 is downloaded and returned as VIDEO. Notably, 提示词 is optional here - the only node in the pack where that's true. Leave it blank and Vidu just interpolates between the frames; add a short prompt if you want to steer the motion (say, "the camera arcs to the left") beyond what the frames imply.

Inputs that matter:

• 起始帧 and 结束帧 - the two IMAGE anchors, both required. Keep their aspect ratios close; the source warns they should stay within about 0.8–1.25 of each other, or Vidu gets confused about the frame it's filling in.
• 运行配置 - the biggest menu in the pack, six options, and the only place you'll see viduq1-classic - 5秒 alongside the usual viduq1, vidu2.0 (4s/8s), and vidu1.5 (4s/8s) profiles.
• 动态幅度 - how much motion Vidu is allowed to invent between your frames. Keep it on 自动 for a straight transition; push to 大 if you want the interpolation to feel alive.

The resolution pre-validation is the same deal as the other video nodes: the viduq1 profiles demand 1080p, and the 8-second profiles demand 720p. Mismatch gets refused before any money moves.

Outputs: video, 封面链接 (cover URL), 任务ID - the standard trio, and the same error behavior (a failure returns the error text on the string outputs and "error" as the task id; the [Vidu::ViduStartEnd2Video] console logs have the real detail).

Install - pack-standard, nothing heavy:

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; ComfyUI Manager search comfyui_vidu_api_node works too. No models to download - the only dependency is requests.

Where people get burned. The aspect-ratio mismatch between your two frames is the most common first-run failure - check that before you blame the API. And remember the pack-wide reality: this is a pay-per-call cloud service, the queue blocks while Vidu renders, you need a ComfyUI build new enough to have the comfy_api module for the VIDEO output, and the profiles are frozen at Vidu Q1/1.5/2.0 from a single July-2025 commit. It's a thin, honest node that does one job well - as long as your two keyframes play nice with each other.