Runway First-Last-Frame to Video

This is the "bookend" generator: you give Runway a starting frame, a finishing frame, and a prompt, and it invents the video that connects them. It's the single most controllable flavor of text-to-video in the Runway family, because the hardest part - what the shot actually looks like - is already decided for you. Your job is just the transition and the motion.

It runs on the Gen3a Turbo model via Runway's API, which is the first thing you should know before you fall in love with it: Runway deprecated Gen3a in mid-2026, and this node is marked deprecated in ComfyUI core. It still works today and it's still in the menu, but it's a sunset model. If you're building something for the long term, check whether Runway's Gen-4 image-to-video node covers your case instead - you lose the end-frame control, but you're on the current model.

How it works

Under the hood it's a straightforward API call. ComfyUI uploads both of your images to its proxy, fires them at Runway's image-to-video endpoint as first and last prompt images, and then polls the task until the finished clip comes back - typically a few minutes, since Runway renders this on their GPUs, not yours. No model files, no VRAM, no local checkpoint. You just need a Comfy account with credits and an internet connection.

The inputs are all the ones you'd guess:

• prompt - what happens between the frames. Describe the motion and the middle of the shot, not the frames themselves; they're already pinned.
• start_frame and end_frame - your two keyframes. The end frame is only supported on Gen3a Turbo, which is what this node always uses, so no surprises there.
• duration - the dropdown Runway offers. The node's own docs point out that complex transitions - where the end frame is nothing like the start frame - benefit from the longer 10-second setting, because it gives the model time to ease between two very different images. Short clips are where first-last-frame gets jumpy.
• ratio - for Gen3a Turbo that's portrait 768:1280 or landscape 1280:768. Square is not an option here.
• seed - set it to reproduce a take; leave 0 for fresh randomness.

Gotchas worth knowing

Your two frames should share an aspect ratio and stay within a 1:2 to 2:1 range - the node validates both, and the API is pickier about mismatched frames than the prompt is. If you're generating the frames yourself in ComfyUI, just make sure they're the same dimensions before you wire them in.

The output is a VIDEO object, so it needs a video save or preview node downstream to actually land on your disk - otherwise the run completes and the clip evaporates. Runway publishes a best-practices guide for keyframes on Gen-3 (linked from the node itself) that's worth a skim before you fight a transition that keeps melting; the short version is: keep the two frames recognizably related in subject, framing and lighting, and let the prompt handle everything in between.

One more honesty note: this is a paid API node, so every run costs credits, and the price badge on the node tracks the duration. It's not the thing to iterate on for fun - it's the thing to use when you already know exactly which two stills your scene needs.