Vidu Start End To Video Generation

Most text-to-video is a leap of faith: you describe the ending and hope the model lands it. This node removes that gamble. You give Vidu a first frame, an end frame, and a prompt, and the model has to produce a video that starts at one and ends at the other. It's the difference between "I hope she ends up looking at the camera" and "she will end up looking at the camera, because that's the last frame you're holding."

It runs the first-generation viduq1 model, which is the fixed, no-frills tier of the Vidu family: 5 seconds, 1080p, done. If you like the result, the Vidu Q3 start/end node is the grown-up version with up to 16 seconds, audio, and two model choices. This one is the simple, cheap way to see whether first-last-frame generation is even how you like to work.

The inputs that matter

• first_frame and end_frame - your two bookend images. This is where all the quality lives, so spend your effort here, not on the prompt. The node quietly checks that the two frames have roughly matching aspect ratios (they need to be within about 80–125% of each other) - if you feed it a portrait start and a landscape end, it'll complain before spending your credits, which is a courtesy.
• prompt - optional, and genuinely optional. The frames are doing the heavy lifting; use the prompt to describe the motion between them ("she walks toward the window") rather than restating what the images already show.
• duration - 5 seconds, fixed. On this node the slider is decorative.
• movement_amplitude - auto, small, medium, large. Your main creative control, and the one people forget. A small setting on a talking-head transition reads as a stable, professional shot; large makes the interpolation do work and can get wild between two very different frames.
• seed - 0 for random, else reproducible takes.

How it runs

ComfyUI uploads both frames to Vidu's start-end endpoint through the Comfy proxy, polls the task, and downloads the finished clip. Needs a Comfy account with credits, internet, no local GPU. Output is a single VIDEO object - save it downstream.

The honest gotchas

First-last-frame generation is only as good as its last frame. If your end frame is a completely different scene from your start frame, the model has to invent a whole transition and that's exactly where Vidu (like every first-last-frame model) starts producing warpy, mushy middle frames. Keep the two frames recognizably related in subject and lighting, and let the prompt describe the bridge rather than trying to bridge two unrelated worlds.

Also worth stating plainly: this is a paid API node, and the cost of a five-second generation is the same whether your frames are great or garbage. Iterate on the stills locally - generate, pick, retry - before you spend a run on them. The community's whole trick with this family is making the expensive part the last thing you do.