Grok Image Edit

Grok's image models are xAI's answer to prompt-driven photo editing - and the answer with the lightest moderation of any mainstream service, which is exactly why a lot of people want them in ComfyUI. GrokImageEditNodeV2 is the current, maintained wrapper for that, the May 2026 rework of the original GrokImageEditNode. The headline change: the model picker is now a DynamicCombo, so selecting a model reconfigures the node's inputs to match what that model supports. Same API under the hood, much less guessing.

Let's get the shared partner-node stuff out of the way in one breath. Nothing runs locally - no VRAM, no checkpoint download. Your reference image(s) are base64'd into the request, posted to Comfy's proxy at api.comfy.org, forwarded to xAI, and the finished image comes back as an IMAGE tensor. Billed from your Comfy account credits; no bring-your-own-key. Seed is explicitly "does the node rerun?" - the tooltip says outright results are nondeterministic regardless of seed, so don't chase reproducibility here.

The model DynamicCombo is the whole game. Each option carries its own nested inputs:

• grok-imagine-image-2.0 - the current flagship. Up to 3 reference images, resolution 1K/2K, number_of_images (1–10), a quality toggle (medium or low), and full aspect-ratio control (yes, even on a single image - the V2 dropped the old "multiple images required" restriction for this model).
• grok-imagine-image-quality / grok-imagine-image - the older tiers, up to 3 references, aspect ratio still needs multiple images on these.
• grok-imagine-image-pro - the premium tier: exactly 1 reference image, no aspect-ratio option, resolution 1K/2K.

The nested inputs are the reason this node feels different from the old one: the reference-image Autogrow slots, resolution, count, quality, and aspect ratio all live inside the model selection, so you only see what the current model supports.

Output is a single IMAGE (batched if you asked for multiple results). Wire it to Save Image or downstream - already decoded, no VAE step.

Practical notes. If you're coming from the original node, the migration is trivial: pick your model in the dropdown, plug images into the nested slot. If your edit involves a single reference and you want a specific frame shape, the 2.0 model is the one - it's the only tier that gives you aspect-ratio control on one image, and it's also just the best current Grok editor. The models are autoregressive vision-token generators rather than diffusion, which shows up as different behavior on fine text and structured detail - sometimes better, sometimes just different, so test against your own subject rather than trusting a benchmark.

Worth it? Grok's edit quality is genuinely strong and the moderation stance is the most permissive of the big API labs, which matters if that's a requirement for your work. The price badge reflects model, resolution, and image count - check it before you set number_of_images to 10 and surprise yourself.