flux2Dev / sdcpp / createImage

Flux 2 Dev is the best open-weights image model most people will never run: 32B parameters with a Mistral-based vision-language text encoder, roughly 50GB of weights end to end, and it wants a 48GB card and patience even then. The community's verdict was never "it's bad" - it was "nobody can hold it." This node is Civitai's answer: the model runs on their sdcpp fleet, you write the prompt, and the image lands back in your graph. No 50GB download, no VRAM budget, just Buzz.

What you're paying for is real. Flux 2 Dev's prompt comprehension and multi-reference composition are a genuine step past everything you can run on a consumer card - it understands complex scenes, handles up to 4MP, and does text-to-image in one checkpoint. For the cost of a cloud job you get the quality ceiling without the hardware. That's the whole pitch, and it's a good one for the occasional "I need the good one" generation.

How it works

The interesting thing about this node in the pack: there's no model input. The recipe fixes the model server-side, so the only required input is prompt. You submit engine: sdcpp, ecosystem: flux2Dev, operation: createImage, the fleet loads Flux 2 Dev, and you get a ComfyUI IMAGE back after a long-poll. No selector wiring, nothing to misconfigure.

The inputs that matter

• prompt (required) - write sentences. Flux 2's Mistral encoder handles long, structured, natural-language prompts dramatically better than CLIP-era models; this is the one node where prompt quality pays off most.
• steps / cfg_scale - defaults 20 / 1. Flux 2 Dev is guidance-distilled, so CFG 1 is the intended setting; raise it only if you're chasing a specific look and know what you're doing.
• width / height - 1024 default, 512–2048 range. It goes to 4MP natively, so upscaling workflows can start from a larger canvas than the Flux.1 nodes allow.
• negative_prompt - exists; Flux 2 barely needs one, and its CFG-1 guidance makes it even less relevant.
• loras - optional, via a Civitai LoRA Selector. There's no large Flux 2 Dev LoRA library yet (training one needs a serious card), but if you have one, this is where it goes.

Install

Pack install: ComfyUI Manager → "Civitai Comfy Nodes" (publisher civitai) → Install → restart. Or git clone https://github.com/civitai/civitai-comfy-nodes.git into custom_nodes, pip install -r civitai-comfy-nodes/requirements.txt (just requests), restart.

Common issues

Credentials are the usual wall: Civitai Auth node into api_config, or CIVITAI_API_TOKEN (from civitai.com/user/account) for headless setups.

The other thing to manage is expectations about cost. This is one of the heavier recipes in the pack - a 32B model running for 20 steps is not a cheap inference job, and it's billed per generation. It's the node you use for a handful of hero images, not for batch exploration. If that's what you need, no local card is competing anyway; if it's not, the Flux.1 or Klein nodes are lighter on the wallet.

Also worth knowing: the whole pack is an official early-preview release and the README warns behavior can change without notice. When a recipe like this shifts, it's the pack, not your graph.