CatsAPI GrokImage

xAI's image model has the shortest on-paper résumé in this pack and the most interesting one. Grok Imagine is the thing that started life as suspiciously Flux-flavored ("Grok on X is using FLUX!" the community shouted, and nobody ever really disproved it) and grew into its own thing - an autoregressive, mixture-of-experts image generator that, as of mid-2026, is one of the least restricted big-name models you can call. CatsAPIGrokImage is the node that reaches it from ComfyUI, one of seven in the ComfyUI_Catsapi pack, all of which are paid-API wrappers over closed models at catsapi.com.

The inputs are refreshingly small

• prompt - multiline text.
• aspect_ratio - eleven presets, 1:1 through 2:1, 20:9, 16:9, 4:3, 3:2, the verticals, and a 1:2 portrait. Notably no size and no quality - Grok Imagine doesn't expose them, and the pack's whole design is to only show what a model actually accepts.
• num_images - 1 to 4 per call.
• max_coins - your spending cap; 0 means unlimited.
• reference_image (optional) - exactly one image. This is the one place the pack is stricter than you'd expect; Grok's API takes a single reference, so don't feed it a batch and expect magic.
• api_key_override (optional) - leave blank locally.

Outputs are the pack standard: images (a genuine IMAGE tensor → Preview Image), file_paths, cost_coins, task_id, metadata. Files land in ComfyUI/output/catsapi/.

Why you'd reach for it

Two reasons. First, if your project bumps into other providers' filters, Grok is historically the least aggressive of the closed image models - the "Spicy Mode" reputation is real even if the exact moderation is opaque and the reseller still passes your prompt to xAI. Don't treat it as an uncensored guarantee; treat it as "the filter loosens up here." Second, it's the cheapest-feeling image node in the pack for plain single-prompt generation - no resolution ladder to climb, just pick an aspect ratio and let the backend do its thing. It's the one I'd grab for a first-pass "what does this look like" check.

How it works

Same skeleton as every node in the pack: preview cost, check max_coins, submit, poll every three seconds until done (15-minute timeout), download with browser headers plus a curl fallback. The result is decoded from a PNG into a local tensor, so you can chain local post-processing. Nothing runs on your GPU, and there's no model to download.

Install and key

cd ComfyUI/custom_nodes
git clone https://github.com/maodeyu180/ComfyUI_Catsapi.git
# restart ComfyUI

Zero extra Python deps. You need a paid CatsAPI key from catsapi.com (starts with cats-):

export CATSAPI_API_KEY=cats-your-key
python main.py

No env var? Drop it in a .env or ~/.catsapi.env; the node reads those too.

Where people get burned

The provider is a Chinese reseller with no meaningful reddit presence - catsapi.com is not an established name, and this pack is brand new, so test with max_coins set low before you commit. Keep the key out of shared workflow JSON (api_key_override serializes into it). And size your expectations on censorship: the looser-filtering claim is vendor marketing until you've seen it yourself, because the underlying model still does its own refusing. One reference image, one aspect ratio, one prompt - that's the whole game, and it's pleasantly boring in the best way.