flux1 / sdcpp / createImage

Flux.1 is the model that made everyone's VRAM cry - dev is a 24GB transformer plus a 5GB T5 text encoder, and Schnell's fast cousin isn't far behind. This node runs it on Civitai's sdcpp cloud fleet instead: you assemble the four-piece Flux stack, type a prompt, and get an image back with zero model downloads. If you've been avoiding Flux because of disk space or an old card, this is the least-pain door in.

The catch that makes this node different from most in the pack: you bring the whole stack. Where SDXL or Anima nodes let you pick one checkpoint, Flux's architecture is four separate pieces - the diffusion model, the VAE, the CLIP-L text encoder, and the T5-XXL text encoder - and this node exposes all four as required CIVITAI_AIR sockets. That's genuinely more setup, but it's also the reason you can mix and match (say, a community Flux dev fine-tune with a different VAE) the same way you would locally.

How it works

Same Orchestration API client as every node in the pack: submit the payload (engine: sdcpp, ecosystem: flux1, operation: createImage) to orchestration.civitai.com, long-poll until done, get a ComfyUI IMAGE back. The four AIR references you wired become the model set the fleet loads for your job. Billed in Buzz, cost shown on the node.

The inputs that matter

• prompt (required) - Flux understands natural language exceptionally well; write sentences, not tag soup.
• diffuser_model, vae_model, clip_lmodel, t5_xxlmodel (all required, CIVITAI_AIR) - wire four Civitai Model Selector nodes (or one, with four outputs if you set it up that way) and pick each piece. The Recipe node has no Browse button - you change models by re-wiring a selector.
• steps / cfg_scale - defaults 28 / 3.5, the classic Flux dev config. Flux is a guidance-robust flow-matching model; CFG above ~4 is rarely useful.
• width / height - 1024 default, capped at 832–1216 in this recipe. Flux trains comfortably at a range of aspect ratios; stay inside the bounds and it behaves.
• negative_prompt - exists and is respected, but the community consensus is that Flux barely needs one; a short one ("blurry, low quality") is all you'd ever lean on.

Install

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

Common issues

Credentials: wire a Civitai Auth node into api_config or set CIVITAI_API_TOKEN (key at civitai.com/user/account); headless installs need the env var.

The Flux-specific failure is an incomplete model stack. A blank or wrong t5_xxlmodel won't throw a friendly error - you'll get a job failure from the orchestrator or a mangled image, because the whole prompt path runs through that encoder. Double-check all four selectors are populated before queueing.

And the pack-level reminder: every run costs Buzz. quantity maxes at 4 here, which at least keeps accidental batches small, but it's still per-image billing. Verify one image before you batch.