Grok Video

The name is a lie, sort of: despite living in your ComfyUI, Grok Video doesn't run anything on your machine. It's one of ComfyUI's built-in partner nodes - a clean front door to xAI's Grok Imagine Video API. You write a prompt (or feed it a starting image), it uploads that to Comfy's API layer, Comfy proxies to xAI, and a VIDEO comes back a minute or two later. No checkpoint file to download, no VRAM involved, no API key to paste. If you have a Comfy account and a few credits, this is the fastest text-to-video you'll ever wire up in a node graph.

This family is how ComfyUI fills the gap for video models it can't run locally: Grok, Kling, Seedance and the rest are closed APIs, and the open alternatives (Wan is the big one, frozen at 2.2 since mid-2025) don't quite match them on polish. The partner nodes are the compromise - you stay in your normal workflow and pay per call instead of renting a GPU.

What you set

• prompt - the text description of the desired video. This does the heavy lifting; be concrete about subject, motion, and camera.
• model - grok-imagine-video or grok-imagine-video-1.5. The 1.5 model is the newer one, and here's the catch: 1.5 always requires an input image. It won't run text-only.
• image (optional) - a starting frame for grok-imagine-video. Required for 1.5.
• resolution - note that 1080p is only offered for the 1.5 model; the older one tops out lower.
• aspect_ratio, duration (1–15s), seed.

About that seed: the tooltip says it plainly - it controls whether the node re-runs, not the result. The actual generation is nondeterministic no matter what you seed. So don't hunt for the "same result twice" behavior you're used to from local sampling; changing the seed just forces a fresh call.

The single VIDEO output feeds anything that accepts video - a Save WEBM node, a preview, or downstream editing nodes like Grok Video Edit.

Where people get burned

• Grok-imagine-video-1.5 + no image is the classic first-run error. If your workflow was text-only and you switched models, that's why it now refuses.
• 1080p on the wrong model throws immediately. Check which model you picked before bumping resolution.
• Cost creep. This is a cloud API billed through your Comfy account credits, priced per call (and resolution/duration scale the price). A "test" that loops a few seeds is real money, so budget like you're spending.

The fix for almost every failure is reading the error string Comfy prints - the node validates resolution, image presence, and empty prompts before anything is billed, so a rejected run shouldn't have cost you anything.