flux2 / klein / createImage

Most of the Civitai image-gen nodes are glorified text boxes with a Buzz meter. CivitaiImageGenFlux2KleinCreateImage is different, and it's the one in this pack that actually feels like a ComfyUI node - because it hands you the controls Comfy people care about: sampler, scheduler, CFG, steps, and LoRAs.

Flux 2 Klein is the size-distilled version of the 32B Flux 2 - 4B and 9B variants, Apache 2.0, the model that reopened the "Flux 2 you can actually run" front when the Dev release's VRAM demands pushed people away. This node runs the hosted Klein, so you get that lineage without downloading anything; your GPU does nothing but display the result. It's part of the official civitai/civitai-comfy-nodes pack (engine flux2, model klein, operation createImage).

How it works. Same machinery as its siblings: build a payload, submit to orchestration.civitai.com, long-poll until done, hand back an IMAGE. The difference is the surface area. cfg_scale (default 5.0, range 1-20) and steps (default 20, 4-50) are the classic dials. sample_method lists the whole sampler menu - euler, euler_a, dpm++2m, ddim_trailing, heun, lcm, tcd, and more - and schedule offers simple, discrete, karras, and friends. Leave them at defaults unless you know why you're changing them, but they're there, which is more than the flex node gives you. There's also a real negative_prompt field, a model_version dropdown (4b, 4b-base, 9b, 9b-base), and loras - a CIVITAI_LORAS socket.

The LoRA part is worth its own paragraph. You don't type a LoRA name into a widget. You wire a Civitai LoRA Selector node's loras output into this node's loras input. The selector holds multiple LoRAs, each with an enable toggle, strength, and a keywords field that's auto-filled from the LoRA's trained words - but those keywords are just a reminder. The LoRA applies by its AIR reference and strength, and to actually invoke the concept you paste the trigger words into this node's prompt. Get that wrong and your LoRA loads but never fires, which is the single most common misunderstanding with this pack.

Inputs you'll actually set. prompt (required), model_version if you want the 9B's extra quality over the faster 4B, and loras if you're stacking styles. width/height (512-2048), seed, quantity (1-4), and enable_prompt_expansion behave as expected.

Install. With the whole pack, via ComfyUI Manager (search Civitai Comfy Nodes) or:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Auth and gotchas. Credentials: a wired Civitai Auth node, then CIVITAI_API_TOKEN, then stored OAuth. 402 = out of Buzz, 401 = bad token, 429 = rate limit. Early-preview pack, so the API-facing shapes can shift between releases. Cloud generation was reported slow at launch, so budget for queue time, and remember every prompt and image leaves your machine to run under Civitai's content rules.

If you want hosted Klein to feel like your local setup - LoRAs, samplers, negatives and all - this is the node. Just remember the keywords-to-prompt dance.