Vidu Q3 Start/End Frame-to-Video Generation

This is the node to reach for when the ending matters. You give Vidu Q3 a first frame, an end frame, and a prompt, and it has to deliver a clip that lands on both - no hope involved. It's the Q3 image-to-video node with a destination pinned, which makes it the natural tool for character scenes where a reaction has to resolve into a specific final look. Combined with Q3's facial-expression strength and the optional audio, it's the closest thing in the Vidu family to "direct a two-beat scene."

Both Q3 models are here - viduq3-pro and viduq3-turbo - with the usual per-model settings nested inside the dropdown (resolution 720p/1080p, duration 1–16 seconds, and the audio toggle for dialogue and sound effects). The pattern should be familiar by now: model first, then its options expand beneath it.

The inputs that matter

• first_frame and end_frame - the bookends. The node validates that their aspect ratios are within about 80–125% of each other, so mismatched frames get caught before they cost you a run. The craft is all here: for a character shot, make the end frame a different expression on the same person and you get a micro-scene - "composed and calm" to "startled" - rather than a morph.
• prompt - max 2000 chars, and for first-last-frame work it should describe the middle. The frames are the punctuation; the prompt is the sentence between them.
• duration - up to 16 seconds on Q3, and longer is genuinely better here: complex transitions between very different frames need the runway, the same lesson Runway's first-last-frame docs teach.
• audio - Q3's signature. On for a scene with dialogue, off while you're composing.
• seed - 1 for reproducible, 0 for random.

How it runs

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

The honest gotchas

First-last-frame generation fails the same way everywhere, and Q3 is no exception: if your two frames are unrelated worlds, the model has to invent a transition and you get warpy middle frames. Keep the subjects, lighting, and general composition recognizably connected, and let the duration and prompt do the bridging. The other real-world note is cost - Q3 pro at long durations with audio is the top of the Vidu price ladder, and the community treats it as a commitment. Iterate the stills locally until you love them, iterate the motion on turbo at 720p, and spend the pro run on the take that ships. Also: moderation is identical to Vidu's website. ComfyUI is the pipeline, not a way around the rules.