flux2Dev / comfy / createImage

Flux 2 Dev is the model that lost the launch and won the year. The 32B rectified-flow transformer - roughly 56B of weights with its Mistral-3 text encoder - is, by broad agreement, the most capable open-weights image model around. It was also rejected in a week because nobody could run it: the KB's math was ~35GB fp8 plus an 18GB encoder. Civitai Image Gen (flux2Dev / comfy / createImage) is the end-run around that: the model runs on Civitai's fleet, so the 56B problem stops being your problem.

What it is

The text-to-image node for the Flux 2 Dev ecosystem through the Comfy engine. Note what's not here: no model socket. Unlike the FLUX.1 node in this pack, the model is fixed - this node always generates on Flux 2 Dev - which removes the Model Selector step entirely. Required input is just prompt.

Inputs that matter

• prompt - required. Flux 2's understanding of complex prompts is its headline feature; the KB quotes people calling it "unparalleled" at composition from multiple references. Write full sentences.
• width / height - 512 to 2048, default 1024. Note the floor: this engine won't go below 512. Flux 2 goes up to ~4MP, so 1536x2048 isn't unusual here.
• steps (4–50, default 20) and cfg_scale (1–20, default 1.0). Flow-matching default of CFG 1 means the model runs guidance-off by default - that's normal for this architecture, not a mistake.
• sampler / scheduler, negative_prompt, loras, quantity (1–4), seed, output_format, image_metadata.

Outputs: images (native IMAGE), errors, workflow_id, raw_json - the standard four.

Why you'd reach for it

The KB's verdict on Flux 2 Dev is that its potential was never hidden, it was gated on hardware: "the best open-source locally available model for image generation in existence" if you own a 48GB+ card and have patience. This node deletes both conditions. For text rendering, complex composition, and reference-heavy prompting, this is the highest-ceiling model in this pack, and the cloud removes the hardware tax that made the community pick Klein over it locally.

That's also the honest caveat: locally, most people chose Klein (the distilled small model) for speed and license reasons. Here you get the big model without the cost of running it locally - you just pay per call in Buzz instead. Whether that's a good trade is entirely about how often you generate.

Install & gotchas

Standard pack install - ComfyUI Manager (search Civitai Comfy Nodes), comfy node registry-install civitai-comfy-nodes, or git clone + pip install -r requirements.txt. Auth via env var / OAuth / Civitai Auth node.

Worth flagging:

1. quantity maxes at 4 here, not 12. The Flux 2 engine caps batches lower than some others. And each image is billed, so quantity is a cost dial as much as a convenience.
2. CFG 1 is the correct default. Bump it if you want, but unlike SD-era habits, Flux 2 doesn't want CFG 4. The 1–20 range with a 1.0 default is the giveaway.
3. No ControlNet socket on this node. The flux2Dev create surface exposes loras but not control_nets - if you need guided structure, this engine's API doesn't take it. Check the sockets before you plan a workflow around it.

The one-line version: this node is the cleanest answer to "I want the best open-weights model's output without the 56B download." It's the argument for API nodes in its purest form - a capability you can never have locally on modest hardware, made available at a metered price.