CatsAPI FLUX.2 Pro

The local FLUX.2 story is one model: the 32B Dev, which needs a 48GB card and patience, or Klein, the tiny distilled line that everyone actually runs. FLUX.2 Pro is the third option - Black Forest Labs' commercial API tier, and it doesn't exist as a download at all. If you want it, you call it, and this node is a tidy way to do exactly that from inside a ComfyUI graph. It's one of seven nodes in the ComfyUI_Catsapi pack, each of which wraps a single closed model behind an API at catsapi.com.

The name is a lie, sort of

"Pro" here means "the hosted, pay-per-image version of FLUX.2," not a bigger local checkpoint. The pack calls it flux2Pro behind the scenes, and it inherits FLUX.2's best trick: multi-reference editing. Wire up to three images into the optional reference_image input and it'll compose from all of them - the thing people say the open FLUX.2 line does better than anything else. You get all of that without renting the GPU to run the 56GB-total model, which is the entire point of an API node.

The inputs, which are refreshingly few

• prompt - your text, multiline. This is the one you'll actually tune.
• aspect_ratio - five options: square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9.
• max_coins - a spending cap in CatsAPI's cat-coins; 0 means no cap. More on that below.
• reference_image (optional) - an IMAGE tensor, up to three images.
• api_key_override (optional) - leave blank locally.

Notice what's not there: no num_images (you get one image per run - the Pro API is single-shot), no size, no quality slider. The node only exposes what the model genuinely accepts, which is the pack's design philosophy and a blessing after a decade of nodes that invent parameters.

Outputs are the pack-standard five: images (a real IMAGE tensor you can plug straight into Preview Image), file_paths (JSON list of where the PNGs landed), cost_coins, task_id, and metadata as a JSON string. The image is downloaded to ComfyUI/output/catsapi/ and converted locally - you can chain a local upscaler or a detailer after it exactly as if a local sampler made it.

How it actually works

Under the hood it's a plain HTTP client: preview the cost, check it against your max_coins, submit the task, poll every three seconds until it's done (15-minute timeout), then download the result with browser-like headers and a curl fallback for CDN bot checks. No model files, no VRAM, no dependencies beyond what ComfyUI already ships.

Install and the key

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

No requirements.txt, no venv dance - the pack uses urllib plus numpy/PIL/torch that ComfyUI already has. You do need a CatsAPI key from catsapi.com, which costs money and starts with cats-:

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

If your launcher doesn't inherit shell env, drop the same line into a .env file or ~/.catsapi.env and the node reads those too. Keys containing ... are rejected as truncated before any request goes out.

Where people get burned

The classic API-node gotchas apply. The provider is a Chinese reseller with zero reddit footprint - catsapi.com isn't kie.ai, and the pack is young (first commit June 2026), so treat it as fresh territory. Don't paste a filled-in api_key_override into a workflow you share, because ComfyUI serializes widget values into the JSON. And remember: your prompt and reference images leave the machine - that's the mechanism, not a bug. If max_coins is set and the preview exceeds it, the node cancels before submitting, which is a small mercy given video-model prices in this pack. FLUX.2 Pro is single-image, so check your spending on drafts, not finished shots.