NS Grok Imagine Image

NS Grok Imagine Image is the text-to-image / image-edit node for xAI's Grok Imagine models, wired into the symbiotica-ai/comfyui-nodes pack. If you want Grok's image output sitting next to your local Flux in the same graph - as a draft, a reference, or a finished asset - this is the node that calls the API and drops the result back as a normal IMAGE tensor.

It's a straight API wrapper, the "closed model, no local weights" case from docs/knowledge/external-api-nodes.md: you cannot download Grok Imagine, so every call costs money and happens in the cloud. What you get in exchange is a frontier image model that needs zero VRAM, zero local download, and produces a result on the same wire format as any local generator - so everything downstream (upscalers, compositors, captioners) just works.

The inputs that matter

• client (required) - a GROK_API_CLIENT from NS Grok Client. No client, no call.
• model - grok-imagine-image-pro, grok-imagine-image (default), or grok-2-image-1212. Pro is the flagship; the base model is cheaper. The older grok-2-image-1212 is there for parity.
• prompt (required, multiline) - the description, or the edit instruction when you supply an image.
• aspect_ratio - 13 options from 1:1 through 16:9, 9:16, and phone ratios like 19.5:9 and 20:9. This is the one to set for vertical formats.
• num_images - 1–10 images per call.
• image_url (optional) - a URL to an image to edit. Empty means pure text-to-image. Note it takes a URL, not a local file path - if your image is on disk, host it somewhere reachable first (or feed the generated result from another node's save).

One output: output_image (IMAGE), ready for any image node.

Install

Part of symbiotica-ai/comfyui-nodes - ComfyUI Manager search "Symbiotica", or clone + pip install -r requirements.txt. No ffmpeg needed here. You need an xAI API key (https://console.x.ai) with billing, wired via NS Grok Client. Restart ComfyUI after installing.

Gotchas

It's metered: pay-per-image, and the Pro model costs more. Set num_images to 1 while you're dialing in a prompt, then widen - nobody wants to pay for four nearly-identical misses. The image_url requirement trips people up more than anything: it's URL-only, so a local path won't work, and a hot-linked URL that blocks xAI's fetch will fail too. And while the API does its own content moderation, treat the model like any API image gen - a refusal comes back as an error string, not a generated image, so check the node's console output when you get nothing back.