Grok Imagine Edit 1-5 Images (xAI)

Why you'd reach for it

Instruction editing is where the ecosystem went in 2026 - instead of scaffolding a character together out of IP-Adapter, PuLID and ControlNet, you hand a model pictures plus a sentence and it does the compositing. This node is that move against xAI's API: up to five source images, one prompt, and you get back a single edited result as a plain IMAGE batch.

The "five" is the reason to pick it over the pile of two-image editing nodes. Person from image 1, garment from image 2, background from image 3, in one request, referenced in the prompt by number. On grok-imagine-image-2.0 that's the supported count; the older models cap at three.

It's also the cheapest form of consistency there is. The author's own tooltip notes that image-2.0 preserves identity, style and layout across repeated edits, which means you can chain Edit nodes for step-by-step work - fix the pose, then the lighting, then the detail - instead of trying to describe the whole final image and hoping. That chained-editing habit is the one thing from the local side of the ecosystem that transfers straight over here.

What's on the node

image_1 is required and it anchors everything: the result keeps its aspect ratio unless you tell it otherwise, and you refer to it in the prompt as "image 1". image_2 through image_5 are optional and stack the same way. Each connected source is billed $0.01 as an input, on top of the output image - so a five-source 2k/medium edit is real money, and the console tells you the estimate before the request goes out.

The prompt is where the work happens. Be concrete and name the sources: "the woman from image 1 wearing the dress from image 2, standing, full body". Say what to keep - "preserve her face and hair" - say what to remove, and add camera or lens language only if it matters. Vague prompts are how you pay twice.

Then the three dials, and they do not overlap:

• aspect_ratio defaults to source, which means the parameter isn't sent at all and image 1's ratio survives. Pick any explicit ratio and the result is re-framed - the model fills the new frame rather than stretching your image into it. That's usually what you want for a crop-to-vertical edit, and rarely what you want by accident.
• resolution is size: source (not sent, currently 1k), 1k (~1 MP) or 2k (~4 MP, +$0.02).
• quality is the detail tier and defaults to medium here - the deliberate opposite of the Generate node, because edits are exactly the job that benefits from more compute per image (skin, fabric, small text). auto, low and default are there if you want them.

n (1–10) gets you several variants of the same edit and multiplies the bill. upload_format and upload_max_side (512–4096) control what actually gets uploaded: PNG is lossless, and the side limit downscales big sources before they go over the wire - worth lowering on a slow connection, worth raising when the whole point of the edit is fine texture.

Outputs

image is the edited batch - feed it back into image_1 of another Edit node for a second pass, or into Grok Image Save to get name_000000.png in a folder you picked. prompt is the passthrough text, so connect it to a save node's prompt_text and your sidecar .txt writes itself. info gives you the truth of the run: model used, real pixel size, how many sources were billed, moderation flag, and estimated vs. actual cost.

Install

ComfyUI Manager, search Grok Imagine, or:

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

Restart, then Ctrl+F5 once. The key comes from the api_key widget, XAI_API_KEY, or grok_api_key.txt in the ComfyUI root / portable root / pack folder - in that order. The same file is shared with the author's comfyui-grok-prompt-forge pack, so you only set it up once.

Where people get burned

The bill. Five inputs plus n: 4 at 2k/medium is not a rounding error, and the estimate prints in the console before dispatch - read it. Moderation is the second one: an edit that comes back subtly wrong is often an edit the API filtered, and info carries the respect_moderation flag, while fail_on_moderation will stop the queue instead of handing you the filtered image. Third: connecting sources you never reference in the prompt. The model isn't psychic about which of your five pictures was supposed to matter, and you paid $0.01 for it anyway.