Grok Reference-to-Video

Grok Video Reference is the "show, don't just tell" node in the Grok video family. Plain Grok Video takes a prompt (or a single starting image); this one lets you hand the model up to seven reference images to anchor style and subject, and - the part that gets people excited - optional preset voice references so the video can include spoken audio in a voice you picked. It's a built-in partner node, so it's a cloud call: upload references, wait, get a VIDEO back. No GPU work on your side, billed through your Comfy account.

The way you steer it is a little unusual and worth reading before you complain that "nothing is happening." With the newer model, you refer to your images by name in the prompt using @Image1 through @Image7 (numbered in input order, and a batched image input counts once per image). Voices work the same way: @Audio1, @Audio2, @Audio3 in the prompt. The model isn't guessing which reference you meant - you have to point at it.

The inputs

• prompt - the text description of the desired video, including the @ImageN / @AudioN tags that tie references into the scene.
• model - a dropdown that carries most of the controls with it:
  • grok-imagine-video-1.5 adds reference_images (up to 7), voice_1/2/3 (preset voices only - the API supports these, not custom audio uploads), plus resolution, aspect_ratio, and duration.
  • grok-imagine-video offers reference images plus the same resolution/aspect/duration knobs, without the voice options.
• seed - the usual partner-node seed: it just triggers a re-run; actual results are nondeterministic.

What comes out

A single VIDEO output. Save it or hand it to another node in your graph.

Gotchas

• Voices are presets, not uploads. The node description and tooltips are explicit: the API supports only the listed preset voices. If you expected to clone your own voice, this node won't do it.
• Un-tagged references are wasted. Feed five images but never mention @Image2 in the prompt and the model has no reason to honor it. Tagging is the mechanism, not an optional extra.
• Watch the reference count. The 1.5 model caps at 7 images; a batched input counts per image. More than that and the request fails.
• It's a paid call per generation, and adding voices/references can push quality and price up together - budget accordingly.

The pattern to steal: generate a character or scene locally, then feed a couple of frames plus a voice into this node to get a video that actually matches what you built, instead of rolling the dice on pure text.