MiniMax H3 First-Last-Frame to Video

Start here, end there, and let MiniMax H3 fill in the middle. That's the entire job of this node: you supply a first frame, optionally a last frame, and it generates the video between them. No aspect ratio to babysit - the video simply follows the shape of your images. For transitions, title sequences, or a hero shot that needs to land on a specific final composition, this is the most controllable mode in the H3 family.

It's a partner node: MiniMax renders it on their servers through Comfy's API proxy, billed per second through your Comfy account. It shipped in ComfyUI core in late July 2026 alongside the rest of the H3 trio. The trick that makes it sing: generate both frames locally (they don't even have to be the same model), then hand them over - you control the endpoints, the cloud handles the physics in between.

How it works

Both images get validated - at least 256 pixels per side, aspect ratio between 0.4:1 and 2.5:1 - then uploaded and sent to MiniMax as the first and last frame of the generation. The model invents the motion between them, guided by the prompt in the model dropdown. With only a first frame, the video drifts freely from your anchor; with both, it commits to arriving at that final frame.

The model dropdown holds the real controls: prompt, resolution (768P or 2K), and duration (4 to 15 seconds). There's no ratio setting here, because the images decide the shape - which is a feature, not a cut-down; it means you can do portrait, landscape, or square just by cropping your frames.

The inputs that matter

• first_frame - the required opening image.
• last_frame - optional; add it when you need the video to land somewhere specific.
• model - prompt, resolution, duration inside the dropdown.
• seed - sent to MiniMax; same seed gives similar, not identical, results.
• watermark - toggles the AIGC watermark, off by default.

Output is a single VIDEO.

Gotchas

The aspect-ratio guard is stricter than you'd expect - images more extreme than 2.5:1 get rejected, so a tall 9:21 vertical or a super-wide banner errors out. And while the motion model is clever, it's not teleporting: a first frame of a city street and a last frame of the same street at night will read as "the light changes", but wildly unrelated frames produce a transition that's more morph than shot. Keep the endpoints in the same world and the middle looks like magic.