Grok Image Edit

xAI's Grok image models are the least-censored mainstream image generators around, and they edit photos from a text prompt - change the subject's expression, restyle the scene, reimagine the whole frame. GrokImageEditNode was the first ComfyUI wrapper for that, added to core in January 2026 when the Grok Imagine nodes landed. Here's the thing you need to know before you even look at the inputs: it's deprecated. Marked is_deprecated, replaced by GrokImageEditNodeV2 (May 2026) which reworked the whole thing around a DynamicCombo model selector and added the newer grok-imagine-image-2.0 model. If your ComfyUI is up to date, use the V2.

But it's worth understanding the original, both because old workflows still reference it and because the underlying mechanics are identical. It's a partner node: nothing runs locally, no VRAM, no model files - your image gets base64-encoded into the request, posted to Comfy's proxy at api.comfy.org, forwarded to xAI's /v1/images/edits endpoint, and the result downloads back as an IMAGE tensor. Billed from your Comfy account credits, no bring-your-own-key.

Inputs on this version: model (a plain dropdown: grok-imagine-image-quality, grok-imagine-image-pro, or grok-imagine-image), image (labeled "images" internally - up to 3 input images, or just 1 for the pro model), prompt, resolution (1K/2K), number_of_images (1–10), seed, and an optional aspect_ratio that only works when multiple images are connected. That last constraint trips people up: the tooltip and source both refuse a custom aspect ratio on a single input image - the API only honors it in multi-image mode. The pro model is also capped at one input image, and the node enforces both limits with clear errors.

Mechanically it's one of the simpler ones - no poll loop, a single edits call that returns one or more finished images (they get concatenated into a batch tensor if you asked for several).

Output is an IMAGE tensor, already decoded. Save it or feed it onward - no VAE step, nothing unusual.

Why the original is worth knowing about beyond history: a lot of Grok edit workflows in the wild are built on the old node, and when you open one you need to recognize what you're looking at. And the Grok models themselves are genuinely distinctive - autoregressive vision-token generation rather than diffusion, which is why they handle text-in-image and fine detail differently from Flux. If you're reusing an old workflow, just swap the node for the V2 and rewire the image input; you get the same behavior plus the newer model and a cleaner interface. If you're starting fresh, don't build on this one - it's the retired version.