EvoLink Grok Imagine Image 2.0 (Official)

Grok Imagine occupies a specific corner of the image-model landscape: it's xAI's image generator, and it's known in the community as the loosest filter of the big closed models - plus it handles multi-image editing unusually well. EvoLink_GrokImagine wraps Grok Imagine Image 2.0 so you can reach it from the ComfyUI canvas without ever opening a browser tab. If your prompt is being stonewalled by the bigger platforms' moderation, this is the model people point at.

It's a do-everything image node: text-to-image, image-to-image, and multi-image editing with a clever <IMAGE_0> reference syntax that lets you point at a specific reference image inside the prompt.

The inputs that matter

• prompt (required) - your description or edit instruction. For multi-image editing, reference the images by index: <IMAGE_0> is the first, <IMAGE_1> the second, and so on. So "make the character in <IMAGE_0> wear a red coat" is a valid edit prompt.
• image1, image2, image3 - reference images. Connect Load Image to image1 and it becomes image-to-image; add more and it becomes multi-image editing. Up to three.
• quality - medium (default) or low. There's no high tier; low is faster and cheaper, good for drafts.
• resolution - 1K (default) or 2K. No 4K here, which is a genuine limit of the model.
• size - 13 aspect ratios including the long cinematic strips (19.5:9, 9:19.5). auto default tends toward portrait.
• n - how many images per run, up to 10. Each is billed independently; note that reference images in edit mode are billed per edit, not multiplied by n.

The usual api_key and timeout_seconds (default 300) close it out.

How it works

Like every node in the pack, the request goes to EvoLink's API (POST /v1/images/generations with grok-imagine-image-2.0 as the model), your local reference images are uploaded through EvoLink's own file channel (auto-deleted after 72 hours, no third-party host), and the pack polls until the image is done. Results come back as an image output you can feed straight into a Save Image or Preview Image node. The other outputs are the pack standard: task_id, status_info, result_urls, response_json.

Installing

No GPU, no model files - only requests:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink.git

Fully restart ComfyUI afterward (process restart, not browser refresh), or install via ComfyUI Manager by searching EvoLink. If the nodes are missing, the usual two causes: a nested folder (the pack's __init__.py must be directly in custom_nodes/comfyui-evolink/) or an incomplete restart.

Troubleshooting

401 key, 402 balance, 403 model access, 429 rate limit - same as the whole pack. The interesting one with Grok is moderation: the "least restrictive filter" is relative, not absolute, and it's still a hosted service enforcing xAI's policy on their servers. If a prompt gets blocked, adjust the wording or the reference image and retry - blocked jobs don't charge. And if the reference syntax isn't working, check that your images are actually wired into the sockets; an empty image input quietly changes the request from edit to pure text-to-image.