Grok Imagine - image to image

What it is

The reference-image mode of xAI's original Grok Imagine image model, under KIE Next / Image / xAI / Grok Imagine. One image in, one prompt, one image out - restyled, re-posed, recontextualised, or just nudged. The catalog documents the input as a single reference image URL, so despite the plural socket name you're feeding it one picture, not a mood board.

Its reputation in this ecosystem is specific: Grok-family image generation is the least restrictive of the big consumer models, which is exactly why people reach for it and exactly why you should read the nsfw_checker tooltip carefully rather than treating the node as a permission slip.

How it works

The connected IMAGE is uploaded to KIE's upload endpoint as PNG and the resulting URL is sent with your prompt to the model's create-task endpoint. The node then polls until the task completes, downloads the result, converts it to a ComfyUI IMAGE tensor, and archives a copy under your output folder. No local model, no VRAM - you're renting this one call at a time.

That upload step is worth remembering for two reasons. It means your reference image leaves your machine, and it means the model receives a re-encoded PNG rather than your original file. For most restyling that's invisible; for work where you care about exact pixel provenance, it isn't.

The inputs

Required:

• image_urls - the IMAGE socket. Wire a LoadImage, a local generation, or anything that emits an image tensor.
• prompt - a description of the result you want, and the docs allow a genuinely enormous string (up to 390,000 characters, which is not a typo). Practically: describe the whole picture, not the delta. "Recreate this movie poster with two anthropomorphic cats in the same pose" is the author's placeholder, and that's the register that works - you're describing the destination, not issuing a repair order.

Optional: nsfw_checker, default false, and timeout_seconds / callback_url (leave the callback empty for normal ComfyUI use).

About nsfw_checker

Read the tooltip in full, because it does the honest thing and undercuts its own promise: set the checker to false and KIE's content filtering is disabled, results come back as the model produced them - and there is no guarantee anything is filtered out, so if you don't like what you get, that's on you. What it is not is a way around the model's own tendencies or xAI's upstream policy. Our KB's read on the closed-model layer holds here: filtering follows the model, not the node, and a reseller can't remove what the vendor enforces. If you need genuinely unrestricted image editing, that's a local checkpoint with a community LoRA, not an API flag.

Outputs

image is the result tensor. url is the primary result URL and all_urls_json is every URL the task returned - raw_json is the full provider response, and credits_consumed / credits_left are floats that also render on the node after a successful run. That little receipt is the most practically useful output in an API pack: it's how you notice a restyling session has cost more than a month of electricity for the local equivalent.

Install

Manager → search KIE.ai Nodes Next, or:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
python -m pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

requests, Pillow, numpy, PyYAML. No model files, no GPU budget, but you do need a funded KIE.ai account and network access. Restart ComfyUI, then paste your API key once under Settings → KIE.ai Nodes Next → Connection; it's validated before saving, stored backend-side, and KIE_API_KEY covers headless installs. The pack builds its nodes from KIE's live docs index on first launch and re-syncs about daily, reloading the frontend once at the start.

Gotchas

• One reference image. The catalog's own description of this field is "up to 1 URL string" - don't expect a multi-reference identity system here. For actual character consistency, the pack's Character Packs are the intended path, and even they warn that a reference plus a prompt is not a universal identity lock.
• Described edits drift. The model redraws; it doesn't composite. Expect the subject to move slightly even when you asked for a background swap.
• Provider cost per call, and no free tier. KIE being a reseller usually means cheaper and easier billing than going direct to xAI - it doesn't mean cheap.
• Moderation is upstream. nsfw_checker=false reduces filtering; it doesn't guarantee a result.
• Missing key surfaces as No KIE API key configured... - the one-time Settings step.