wan / v2.1 / civitai

Wan 2.1 is the Apache 2.0 Alibaba model that made open-source video genuinely good - and this node runs it without you downloading a byte of it. It's the "civitai" provider variant in the civitai-comfy-nodes pack: the generation happens on Civitai's own fleet, billed in Buzz, and the finished clip returns as a native VIDEO. It's also the most "local ComfyUI, minus the hardware" node in the Wan family, because it exposes the knobs local Wan users actually turn.

The interesting bit is the model socket. Unlike most video nodes here, Wan 2.1 on the Civitai provider takes a CIVITAI_AIR model reference - you pick a Wan 2.1 checkpoint from the Civitai library with the Civitai Model Selector node, and Civitai runs that model on their fleet. So you're not locked to a single engine default; the site's own Wan 2.1 checkpoints and LoRAs apply.

The inputs that matter

prompt and cfg_scale (default 4) are required. Then:

• width / height - default 480×480. This is the local-ComfyUI feel: you set the resolution directly instead of picking "720p." Wan 2.1's comfort zone is around 480-720p; don't crank past what the model was trained for and expect magic.
• model - the CIVITAI_AIR socket. Wire the Civitai Model Selector's air output in to choose which Wan 2.1 checkpoint runs. Empty = platform default.
• loras - a CIVITAI_LORAS socket from the Civitai LoRA Selector. Stackable Wan 2.1 LoRAs with per-LoRA strength.
• source_image / images - image inputs for image-to-video.
• steps - 10 to 50, default 20. More steps, more fidelity, slower queue.
• duration - 1 to 30 seconds (default 5); frame_rate - default 24.
• seed, api_config - the usuals.

A note on LoRAs: the Civitai LoRA Selector's keywords field is auto-filled purely as a reminder - to actually invoke the trained concept, paste those trigger words into this node's prompt. The generator applies the LoRA by its AIR reference plus strength, and ignores the keywords field otherwise.

Outputs

Standard pack trio: video (VIDEO) into a Save Video / preview node, workflow_id (STRING) for sidebar tracking, raw_json (STRING) for the response and per-run cost.

Install and auth

Shared one-time setup. ComfyUI Manager → search Civitai Comfy Nodes → install, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt   # just `requests`

Auth: Civitai Auth node, CIVITAI_API_TOKEN, or sidebar sign-in. Headless: env var.

Troubleshooting

• The model socket won't fill. It's not a dropdown - it's a typed socket. You must wire the Civitai Model Selector (or a selector's air output) into it; recipe nodes deliberately have no Browse button of their own. Same story for loras and the Civitai LoRA Selector.
• Blurry high-res output? Wan 2.1 480p-era weights don't get sharper because you asked for 2048px. Keep width/height in the model's range and upscale the returned clip downstream instead.
• Slow / timeout. This pack launched June 2026 during Civitai's GPU crunch; long queues were the norm. Default timeout is 30 minutes - raise via Civitai Auth's timeout_minutes or CIVITAI_COMFY_TIMEOUT.
• Why pick Wan 2.1 when 2.2 exists? 2.1 is older, but it's the fully-open, well-understood base with a huge LoRA ecosystem - and it's the one you can bring your own model to on this provider. If you don't need BYO-model, the fal-hosted Wan 2.1/2.2 nodes are the simpler path.

Reach for this node when you want the classic local-Wan workflow - custom checkpoint, LoRAs, exact resolution - but would rather rent Civitai's GPUs than buy one.