flux1 / comfy / createImage

FLUX.1 was the model that replaced SDXL as the open-weights quality bar in 2024 - the 12B transformer with the T5 encoder that finally made natural-language prompting work. Civitai Image Gen (flux1 / comfy / createImage) lets you generate with FLUX.1 checkpoints on Civitai's cloud, with no 12B stack to quantize and no VRAM math. You pick the checkpoint, type the prompt, and the fleet does the rest.

What it is

One of the pack's per-engine image nodes. The display name reads like a file path - flux1 / comfy / createImage - because the ~160 generated nodes are organized ecosystem-first, and this one is the text-to-image operation for the FLUX.1 ecosystem through the Comfy engine.

Unlike the Flux 2 node in this pack, this one has a required model input (type CIVITAI_AIR): you fill it by wiring a Civitai Model Selector (its 🔍 Browse Civitai button shows generation-capable models). That's the whole loop - the model is referenced by AIR URN, nothing downloads locally, and the job runs on Civitai's fleet for Buzz.

Inputs that matter

• prompt - required, natural language. FLUX's T5 encoder reads full sentences, so write like a person, not a tag soup (that's the FLUX lesson from the KB's flux ecosystem doc).
• model - required AIR socket from a Model Selector.
• width / height (64–2048, default 1024), steps (default 20), cfg_scale (default 3.5) - the basics.
• sampler / scheduler - the menus mirror ComfyUI's own (euler, euler_ancestral, dpm_2, karras, sgm_uniform...). Leave them empty to take the engine's defaults until you have a reason.
• quantity (1–12) - how many images in one job.
• loras - a CIVITAI_LORAS socket from the pack's LoRA Selector.
• vae_model - optional AIR override if you need a specific VAE.
• control_nets - wire a Civitai ControlNet stack for edge/depth/pose conditioning.

Outputs: images (native IMAGE, ready to feed any downstream Comfy node), plus errors, workflow_id, and raw_json for debugging and cost.

Install & gotchas

Standard pack install - ComfyUI Manager (search Civitai Comfy Nodes), comfy node registry-install civitai-comfy-nodes, or git clone. Auth: env var, stored OAuth, or a Civitai Auth node.

The things that actually trip people:

1. The required model socket is the usual first hang-up. You must wire a Model Selector in; there's no Browse button on the recipe node itself (by design - the README says change a model by rewiring the selector). A fresh, un-wired node refuses to queue.
2. Know which FLUX.1 you're picking. Dev vs Schnell matter - the KB's honest mid-2026 take is that FLUX.1 is a legacy model with a surviving niche rather than a default, and Dev's license is non-commercial. The Selector shows thumbnails; check the model card before you spend Buzz.
3. It's metered, per image. quantity 12 is 12 billed images. The raw_json/workflow_id outputs are there precisely so you can see what a run cost.

One take: if you want FLUX-quality output and you don't have a 24GB card, this node is the honest route - the same way the KB frames API nodes generally: the model isn't on your machine, so you call it. If you do have the hardware, local FLUX.1 via GGUF/fp8 is free per call and this node's appeal shrinks fast.