Grok Imagine Generate (xAI)

Why you'd reach for it

Grok Imagine has no weights. You can't download it, you can't quantize it, you can't run it on a 3090 - the node is an HTTP client with your xAI key attached, posting to api.x.ai/v1/images/generations and dropping the result back into your graph as an ordinary IMAGE batch. From the canvas it looks like any other generator.

The reason people bother is price. grok-imagine-image-2.0 at 1k on the low tier is $0.04 an image, and the read going around during 2026 was that it lands embarrassingly close to the expensive closed models for a tenth of the money. In one +240 thread about Qwen Image 2.0 staying closed, the argument that kept resurfacing was simply: why would anyone bother, when grok-imagine-image is $0.02 and isn't that far off Nano Banana. That's the whole case. Cheap closed-model quality, inside the same graph as your upscaler and your masks.

The cost is the usual one for API nodes: prompts leave your machine and the provider's filter applies, with nothing local to patch around it. Whether that's a dealbreaker is your call, not the node's.

The inputs you actually touch

prompt is the whole job. image-2.0 follows plain sentences closely - name the subject, action, setting, camera, light, style - and it's unusually good at text, so put any lettering in quotes and say where it goes. There's no negative prompt field, because the API has no concept of one. Say what you want instead of what you don't.

Three widgets get people in trouble, and they're worth understanding before you spend anything:

• resolution is size, and there are exactly two classes: 1k (~1 MP) and 2k (~4 MP). 2k costs $0.02 more per image.
• quality is detail tier, not size: low, medium, auto, default. The author defaults it to low so the billed tier is visible instead of hidden behind auto. Medium is roughly double the price. The classic mistake is cranking resolution and wondering why the image is still softish - that's the other dial.
• n (1–10) generates the batch in one request. Every image is billed, so n: 10 at 2k/medium is $0.80 in one click. The console prints the estimate before it sends anything.

seed deserves a sentence of its own: it is not sent to the API and results are not reproducible. It's a cache-breaker. Leave it fixed and an unchanged graph reuses Comfy's cached output and costs nothing; randomize it (or change it by hand) and you get a fresh, freshly-billed request. fail_on_moderation turns a filtered result into a hard queue error instead of a silently degraded image - useful when you'd rather stop than ship something the API declined. custom_model overrides the dropdown if xAI ships a new ID before the pack updates.

Worth knowing: quality is only sent when the model starts with grok-imagine-image-2.0. On the older grok-imagine-image it's ignored, which is correct, not broken.

Outputs and where they go

images is the batch - wire it into an upscaler, a Preview, or Grok Image Save if you want numbered files in a folder you chose instead of the ComfyUI output directory. prompt is the text exactly as sent, which is what you connect to a save node's prompt_text so your .txt sidecars are never empty. info is one line of facts: model actually used, true pixel size of each image, requested resolution and quality, the moderation flag, and estimated vs. billed cost in USD. When a bill surprises you, info is where you find out why.

Install

ComfyUI Manager, search Grok Imagine, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Slartibart23/comfyui-grok-imagine-video
..\..\python_embeded\python.exe -m pip install -r comfyui-grok-imagine-video\requirements.txt

(Portable Windows layout above; on Linux/mac use your own python against the same requirements file.) Then restart ComfyUI and hit Ctrl+F5 in the browser once - that loads the pack's preview script. Dependencies are light: requests, numpy, Pillow, imageio, imageio-ffmpeg. No models to download, no VRAM used.

The key is found in this order: the api_key widget, then XAI_API_KEY, then grok_api_key.txt sitting in the ComfyUI root, the portable root, or the pack folder. Use the env var or the file - the widget value gets baked into any workflow you export.

When it goes sideways

No xAI API key found means all three lookups missed; get a key at console.x.ai and drop it in ComfyUI/grok_api_key.txt. HTTP 429 and the 5xx range are retried automatically a couple of times with backoff before failing, so a single blip usually isn't your problem. grok-imagine-image-quality is retired on 2026-11-02 and the node warns if it's still selected - after that it's served by 2.0 at low quality, so just switch to grok-imagine-image-2.0. And if a result comes back visibly blander than your prompt, check info for the moderation flag before blaming the model.