sd1 / comfy / createImage

Let's not sugarcoat it: SD1.5 is the model you can run on a potato. A 512×512 generation takes a few seconds on any half-decent card, the checkpoints are a couple of gigabytes, and the ecosystem around it is the most documented in the entire hobby. So a node that sends your prompt to Civitai's cloud, runs SD1.5 on their GPU, and bills you in Buzz for the privilege is, on paper, the definition of the wrong tool - the community's own frame for API nodes is "right for a model you can't run, wrong for one you can."

And yet. There are three reasons this node exists and gets used. First, you might not have a GPU at all - a laptop, a work machine, a cloud VM that never got a card. Second, the Civitai Model Selector integration is genuinely nice: wire a selector into the required model socket and you can browse Civitai's SD1.5 catalog in a card grid and switch checkpoints by rewiring, no file downloads. Third, consistency: everyone in the same team runs the exact same hosted pipeline, which removes the "works on my machine" variance of local installs. If any of those apply to you, this is the node. If none do, you already know the local route is free and faster.

The inputs that matter

This one has two required inputs: prompt and model (a CIVITAI_AIR socket - you must wire a Civitai Model Selector into it; there's no built-in default checkpoint).

• negative_prompt - SD1.5 lives and dies by it. Use it.
• steps (30) / cfg_scale (7) - the SD1.5 classics. 30 steps at cfg 7 is right where most SD1.5 work has always sat.
• width / height - 512×512 default. SD1.5 does not want to go big; that's what the SDXL node is for.
• clip_skip - default 2, which is the community's preferred setting for most SD1.5 checkpoints (skips the last CLIP layer for richer embeddings). Nice default, and it's a socket-free widget so it just works.
• sampler / scheduler - the standard ComfyUI menus. If you have a favorite, set it.
• vae_model / loras / control_nets - CIVITAI_AIR, CIVITAI_LORAS, and CIVITAI_CONTROLNETS sockets, filled from the Civitai Selector nodes. The ControlNet input is the standout: a hosted SD1.5 ControlNet pass without a local ControlNet install.
• quantity, seed, output_format, image_metadata - standard.

Outputs: images (IMAGE), errors, workflow_id, raw_json.

Install and auth

Same pack-wide setup. ComfyUI Manager → 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

Authenticate with a Civitai account with Buzz: export CIVITAI_API_TOKEN=..., a Civitai Auth node, or the sidebar sign-in.

The honest call

The most common failure here is wiring the selector wrong - remember the model socket is required, so an empty graph will fail at run time with a clear "missing required input" error before anything is billed. Otherwise the biggest real-world gotcha is expectation-setting: you're paying per call for a model you could run for the cost of electricity. If you're only on this node because your machine can't run SD1.5, fine - but check the SDXL node's 1024×1024 output before committing, because the modern Civitai catalog is built around it. And keep the pack's early-preview warning in mind: Civitai regenerates these nodes from their API spec and says behavior may change without notice.