🔹 Now in early preview — installable from the Comfy Registry. Still under active development: nodes, APIs, and behavior may change without notice. Use at your own risk.

ComfyUI custom nodes for the Civitai Orchestration API. Run Civitai's cloud recipes — image/video/audio generation, upscaling, training, captioning, moderation — as nodes inside any local ComfyUI graph. No local GPU or model downloads needed; jobs run on Civitai's fleet and are billed in Buzz.

Install

From the Comfy Registry (recommended)

The pack is published to the Comfy Registry, so ComfyUI Manager can install and update it for you:

In ComfyUI Manager: open Manager → Custom Nodes Manager, search for Civitai Comfy Nodes (publisher civitai), and click Install, then restart ComfyUI.

With comfy-cli:

comfy node registry-install civitai-comfy-nodes

See the Comfy Registry docs for details.

From source

Clone (or unzip) into your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt   # `requests` + `python-socketio[client]`

Install the requirements with the Python that runs ComfyUI, not whatever pip is on your PATH. On the Windows portable build that is the embedded interpreter:

C:\path\to\ComfyUI_windows_portable\python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\civitai-comfy-nodes\requirements.txt

For a venv-based install, activate the venv first (or use <venv>/bin/python -m pip …).

Authentication

Nodes resolve credentials in this order:

A connected Civitai Auth node (explicit token / base URL / mature-content / timeout overrides)
The CIVITAI_API_TOKEN environment variable (create an API key)
A stored API key (~/.civitai/comfy-api-key) or OAuth login (~/.civitai/comfy-oauth.json, auto-refreshed) — both can be set from the Civitai sidebar's connect panel (no env var needed)
An interactive browser login (OAuth + PKCE) — requires CIVITAI_OAUTH_CLIENT_ID to be configured

Headless/remote ComfyUI installs should use the env var.

Nodes

~160 nodes under the Civitai category, generated from the orchestration OpenAPI spec. The menu is organized ecosystem-first — Civitai/<media>/<ecosystem>[/<engine>]/… — with the engine (sdcpp/comfy) shown as a sub-level only when an ecosystem is reachable through more than one engine (e.g. Civitai/Image/zImage › zImage / turbo / createImage, Civitai/Image/anima/sdcpp). Each discriminator variant is its own node so it shows only the inputs that variant actually uses:

Civitai/Image — Image Gen (one node per engine: Flux2, OpenAI, Google, Seedream, …), Upscaler, Background Removal, Image To SVG
Civitai/Video — Video Gen (one node per engine: Wan, Kling, Vidu, Veo3, LTX, Sora, …), Upscaler, Interpolation, Enhancement, Background Removal
Civitai/Audio — Text To Speech, Transcription, Audio Captioning, ACE Step Audio, MiniMax Music 3
Civitai/Text — Chat Completion (plus a simple single-turn wrapper), Prompt Enhancement, Media Captioning
Civitai/Analysis — Media Rating, WD Tagging, XGuard Moderation, Shieldstral Moderation
Civitai/Training — Training, Image Resource Training
Civitai/Misc — Poly Gen (3D mesh generation), 3D Model Preview
Civitai/Loaders — Model Selector, LoRA Selector, Embedding Selector, ControlNet (see below)

Every node returns its media outputs as native Comfy types (IMAGE/VIDEO/AUDIO) plus workflow_id and raw_json for debugging and cost inspection. Models and LoRAs are referenced by AIR URNs (e.g. urn:air:sdxl:checkpoint:civitai:101055@128078).

Models, LoRAs, ControlNets & embeddings

Recipe nodes expose their model references as typed sockets, not text widgets — model/vae are CIVITAI_AIR, loras is CIVITAI_LORAS, embeddings is CIVITAI_EMBEDDINGS. You fill them by wiring a Civitai/Loaders selector node; each selector has a 🔍 Browse Civitai button (a searchable card grid of generation-capable models via a same-origin proxy to civitai.com/api/v1/models) and shows an on-node preview (thumbnail + name) of its current resource. Recipe nodes themselves have no Browse button — change a model by wiring a selector.

Civitai Model Selector — pick a model once; it serves two purposes through its two outputs:
1. Choose the model a Civitai recipe node runs on. Wire its air output into a recipe node's model / vae socket. The job runs on Civitai's fleet, so nothing is downloaded locally.
2. Auto-download a model for a local loader. Wire its path output into any standard loader's file widget — e.g. drop it in front of a Load LoRA (lora_name) or Load Checkpoint (ckpt_name) node and it downloads the model into the matching ComfyUI folder for you, no manual file management. No need to replace the loader — just feed it.
The download happens only when path is connected, so use (1) stays cloud-only and never pulls files down.
Civitai LoRA Selector — holds multiple LoRAs in one node: each row has an enable toggle, the model (pick/replace via Browse Civitai), a strength, and a keywords field; ＋ Add LoRA appends another. Wire its loras output into a recipe node's loras / additional_networks input (chain another selector via the loras input to combine stacks). Wire MODEL + CLIP to also download & apply the enabled LoRAs locally. The keywords field is auto-filled from the LoRA's trained words purely as a reminder — a LoRA applies by its AIR + strength, so the generator ignores these words; to actually invoke the trained concept, paste them into your recipe node's prompt.
Civitai Embedding Selector — pick textual-inversion embeddings; chain (embeddings → embeddings) and wire into a recipe node's embeddings input. Embeddings apply automatically (the sdcpp pipeline prepends each one to the positive prompt), so no trigger word is needed. To use a negative embedding, reference it by its model name in the recipe node's negative prompt — that places it on the negative side instead. (Only sdcpp ecosystems — sd1/sdxl — expose an embeddings input.)
Civitai ControlNet — pick a preprocessor, weight, step range, optional control image; chain and wire into a recipe node's control_nets input.

chat messages and other freeform structures remain JSON text inputs.

Browse your generations

The Civitai sidebar tab (the logo icon in the left rail) lists your Civitai generation history — every workflow you've run, across all media types and sources (web / API / ComfyUI) — pulled from the orchestrator and scoped to your account. Filter by media kind, scroll to paginate, and click any result for a lightbox. Pull a result back into the graph three ways: add to canvas (creates the matching loader node — image→LoadImage, video/audio/3D→their loaders — wired to the imported file), fill a selected loader node, or drag a thumbnail onto the canvas. If no credentials are configured, the tab shows a connect panel (OAuth sign-in or paste an API key).

Civitai Link

The pack is a Civitai Link client: pair it once and the download via Civitai Link button on any civitai.com model page downloads that model straight into the matching ComfyUI folder (models/loras, models/checkpoints, models/diffusion_models, …), verifies its SHA256, and refreshes the model lists. The site also sees which models you already have, and the Model Library shows a Civitai icon next to every model it recognises (click to open it).

Open the Civitai sidebar tab and click Connect with Civitai. The sign-in that opens in your browser also covers Link, so this ComfyUI pairs itself as soon as you approve. The instance shows up under Civitai Link on civitai.com, where it can be renamed or revoked; if you were already signed in, the panel offers a one-click Pair this ComfyUI instead.

If ComfyUI runs on another machine (the browser can't reach it) or you connected with an API key, use a pairing code instead: on civitai.com open Civitai Link (the link icon in the header), add an instance, copy its 6-character code and paste it into the panel. A code pairing can be switched to the account pairing later from the same panel.

The pairing is stored in ~/.civitai/comfy-link.json and reconnects automatically after restarts; ~/.civitai/comfy-install-id identifies this install so re-pairing replaces the same instance instead of adding one (civitai.com allows ten per account). Removing a model from the site's Link panel deletes the file locally. Toggle the feature under Settings › Civitai › Civitai Link; a CIVITAI_LINK_URL env var (or the settings entry) points at a different relay. Link stays off in hosted comfy-cloud sessions.

Development

Nodes are generated — never edit civitai_comfy_nodes/generated/ by hand. To change node shapes, edit codegen/overrides.json (or the codegen itself) and regenerate:

python -m codegen.generate        # regenerate from spec/v2-consumers.json
pytest tests -q                   # unit tests (no ComfyUI or network needed)
CIVITAI_API_TOKEN=... pytest -m e2e -o addopts="" tests/test_e2e.py   # prod smoke test

To pick up orchestration API changes, see scripts/sync-spec.sh.