wan / v2.2 / fal

Wan 2.2 is the last open Wan, and also the best one for still images - the community's verdict was basically "better than Flux, plus it's Apache 2.0." It's a video model, but its still-frame quality is genuinely elite, because training on video teaches a model how 3D space and motion work. So this node is the odd case: an open model you could run locally (the 14B MoE fits on a big card with offloading), wrapped as a cloud recipe that runs on the fal engine behind Civitai's Orchestration API. You're renting someone's GPU for a model whose weights are free.

Which is fine if you have no GPU, or want a quick still without babysitting a two-pass Wan pipeline. The 14B two-pass architecture (separate high-noise and low-noise models) is fiddly to set up locally; here it's one node with sane defaults.

Inputs and outputs

Only prompt is required. The defaults are where this node earns its keep:

• steps - 27 default (the 14B full model likes more steps than the 5B), vs. 40 on the v2.2-5b node.
• acceleration - none / fast / faster, default none. This maps to the community's speed-LoRA debate: speedups cut generation time but the loud consensus was that aggressive acceleration "kills everything that's good" about 2.2 - composition, lighting, emotion. none is the quality setting; reach for fast only when you're prototyping.
• guidance_scale - 3.5 default.
• image_size - a string (square_hd default), the model's preset canvas choices rather than free width/height.
• enable_prompt_expansion, enable_safety_checker - both off by default; flip prompt expansion on if you want the model to flesh out your prompt.
• loras - the selector stack; Wan has a real LoRA ecosystem.

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

Install

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

Or ComfyUI Manager → "Civitai Comfy Nodes" → Install → restart. Authenticate with CIVITAI_API_TOKEN (headless) or the sidebar connect panel; jobs bill in Buzz.

Gotchas

The two-pass architecture means each job is heavier than a single-model sampler, and it's running on a shared queue - slow is normal, and the pack's launch feedback was blunt about it ("generation times are super slow"). If a big canvas times out, retry or drop to the 5B variant.

The honest math: Wan 2.2 is the one model in this pack where the "you can't run it locally" justification is weakest. If you have a 12GB+ card and a ComfyUI install, you can run 2.2-5B today with ComfyUI's native support and pay electricity instead of Buzz. This node is for when that's not true for you. But also: it's the last open Wan - everything after 2.2 is closed API-only, which is exactly the territory the v2.5 and v2.7 nodes in this same pack occupy. Keep 2.2 as the open reference point; it's why the "last open model" keeps showing up in conversations about what comes next.