Runware Reference Videos

Reference-based video gen is exactly the job that makes people start shopping for a second GPU. This node hands that job to Runware's cloud instead. It's a builder node in the Runware/Params group: you tell it which clip to use and how, it hands back a typed referenceVideos value, and you wire that into a video model that accepts it - Wan 2.6, Wan 2.7, SkyReels V4, the Kling video nodes, Seedance 2.0. No local GPU, no model files, no VRAM math. Your machine just draws the wires and pays for compute.

How it works

Everything in ComfyUI-Runware runs through Runware's REST API, and the builder nodes exist to keep the model nodes clean. RunwareBuild_referenceVideos produces a RUNWARE_REFERENCEVIDEOS object that the model node tucks into its referenceVideos request parameter. It's a stackable builder, which in this pack means you can chain multiple reference-video nodes into one socket and send several clips in a single request.

The three inputs that matter

The whole node is three required fields, and each one is doing something specific:

• type - reference or extend. reference tells the model to use the clip as a motion or style guide; extend asks it to continue from the end of that clip. This is the switch that decides whether you're imitating or continuing.
• tag - an identifier like @video1 that links the clip to a mention in your prompt. This is how the model knows which reference applies where. If you tag the clip @video1, your positive prompt has to actually say @video1 somewhere, or the reference never gets attached.
• video - the clip itself, as a URL, path, or Runware media UUID. The constraint that trips people up: max 15 seconds, and a minimum of 1,866,240 total pixels (the tooltip's example is 1920×972). A 1280×720 clip is below that floor, so upscale or crop before you feed it in.

There's no optional config here. That's it - reference the clip, name it, wire it in.

Wiring it in

The output is referenceVideos, and it only plugs into a model node that exposes that socket. In the current catalog that's the video generation crowd - Wan 2.6/2.7, SkyReels V4, the Kling video nodes, Seedance 2.0. Drop the model node, feed it your prompt, wire the builder's output into its referenceVideos input, and mention the tag in the prompt.

Install and setup

This pack is one of the cleanest installs in ComfyUI because it downloads nothing heavy - the requirements.txt is just runware-sdk, pillow, and soundfile. Install via ComfyUI Manager (Custom Nodes Manager → search "Runware" → install → restart), or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart ComfyUI. You'll need a Runware API key: paste it in ComfyUI Settings → Runware API key, set RUNWARE_API_KEY, or run runware auth login once. The key isn't optional - the pack's whole premise is that inference happens on Runware's machines. Account funding is the friction most people hit: Runware's minimum top-up is around $20 with a card on file, which reads as steep if you only want to test once.

Troubleshooting

• Clip rejected - almost always the pixel floor. Below 1,866,240 total pixels the API refuses it. Upscale first.
• Nothing happens in the output - check that your tag (@video1 etc.) actually appears in the prompt. A tag that's never mentioned is a reference that never attaches.
• Audio is a separate problem - reference videos drive motion/structure; if the model needs the clip's audio too, that's a different input on the model node, not this one.

It's a tiny node, but it's the difference between "I describe a camera move" and "I show the model a clip and it does that." For anyone running video work on a laptop, that trade is worth a few cents a run.