Grok Imagine Video 1.5

Grok Imagine is xAI's image-and-video generation family - the thing behind the Grok app's "imagine" feature - and this node exposes the video side (xai:[email protected]) as a plain ComfyUI node. Its defining input isn't a prompt, it's frames: frameImages is required, so this is an image-to-video node at heart. You give it a starting image (or a few frames), describe the motion, and it animates the scene.

If you've ever generated a still you love and thought "now what does this look like moving," this is the node. It's a closed, API-only model, so there's no weights to download and no local run possible - Runware is one of the ways it exists inside ComfyUI at all. Install the pack via ComfyUI Manager (search Runware) or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes/, add a Runware API key, and the whole catalog - this included - appears in the node menu.

How it works

Frame in, video out. The model takes the visual content of your frameImages and extrapolates motion from it according to the prompt. That's a different job from text-to-video: the identity of the subject is already fixed by your input, so the prompt's job is movement - camera moves, physics, what happens next - not appearance. Feed it a still of a city street and prompt "rain starts, camera tilts down"; feed it a portrait and prompt "slow smile, wind in hair."

Inputs that matter

• frameImages (required) - your still or frames, as an IMAGE socket. This is what makes the model worth using.
• positivePrompt (required) - the motion, not a full scene description.
• size - 25 presets, default 480p (16:9). Start small; you can go up once the motion works.
• duration - 1–15 seconds. The model's fps × duration gives your frame count.
• numberResults (1–4) - variations with different seeds.
• safety gate (checkContent, safety.mode dropdown) - the standard set.

The single video (VIDEO) output feeds SaveVideo - the pack's example video workflow is literally model → SaveVideo, and it doesn't get simpler than that.

Where people trip up

• Prompting it like text-to-video. If you write a full scene description, the model has to reconcile it with a frame that already shows something else. Keep prompts to motion and camera.
• The 480p default is a feature. A 15-second clip at high res is a real bill. Draft at 480p, promote only the winners.
• Cloud round-trip. You're waiting on xAI's hardware, so a clip is seconds-to-minutes of queue. Plan around it.
• It wants a strong frame. A mushy, low-detail input gives mushy motion - the model animates what it can see.

The ctr on this page (15%+ of searchers click) tells the story: people are curious about Grok's video model. It delivers when the frame is good and the prompt is about motion. That's the whole skill, and it's learnable in a few runs.