Runway Gen3 Image-to-Video @fal

Feed it a still image and a text prompt, and it animates that image into a short video using Runway Gen3 - one of the better commercial image-to-video models. The important thing to get straight up front: this node does not run anything on your GPU. It's an API call. Runway's model runs on fal.ai's servers, you pay fal for the generation, and the node hands you back a URL. Your 3090 sits idle the whole time.

What it actually is

Runway is one of the closed, hosted video generators (alongside Kling, Luma, Pika). You can't download Gen3 and run it locally - it lives behind an API. fal.ai is a service that resells access to a pile of these hosted models with one key and one billing account, and this node is a thin client for fal's Runway Gen3 image-to-video endpoint. So it's a bridge: keep your compositing and prep in ComfyUI, outsource the one step (the video model) you can't run yourself.

That also means the output isn't frames. It's a STRING - the URL of the finished video on fal's storage. To actually see it in your graph, you load it back with something like LoadVideoFromURL (also in this pack). That two-node pairing is the intended flow.

The inputs

• image - the still you want animated. Required.
• prompt - text describing the motion/scene you want. Runway leans on this to decide what moves and how.
• duration - 5 or 10 seconds. Longer costs more.
• aspect_ratio - 16:9 or 9:16 (landscape or vertical).
• fal_key - your fal.ai API key. Non-negotiable; without it there's nothing to call.

Output: STRING, the video URL.

Installing it

Get the pack via ComfyUI Manager (search mixlab, install comfyui-mixlab-nodes, restart) or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/shadowcz007/comfyui-mixlab-nodes

then install requirements and restart. There's no model to download for this node - the "model" lives on fal's servers. What you need instead is a fal.ai account with credit and an API key from your fal dashboard. Paste it into fal_key.

Things to know before you burn credits

• This costs real money per run, billed by fal, not ComfyUI. A 10-second clip costs more than a 5. There's no free local fallback - if that's a dealbreaker, an open model like LTX-Video or Wan runs on your own GPU instead.
• The output is a link, not a clip. Don't wire the STRING into a video-save node expecting frames. Route it into LoadVideoFromURL (or just copy the URL) to bring the actual video into your graph.
• Bad or missing key = instant failure. If the node errors immediately without spending time, check that fal_key is filled in and valid, and that your fal account has credit.
• Network required. Obvious, but worth stating: this node needs outbound internet. On an air-gapped or locked-down box it simply can't work.
• Prompt matters more than you'd think. Gen3 is prompt-driven for motion. A vague prompt gives you a barely-moving image; describe the action you want.

One last aside - fal periodically revises which model versions and endpoints it hosts, so if a call starts failing after months of working, it's worth checking that fal still exposes this exact Runway Gen3 endpoint rather than assuming the node broke.