wan / v2.2-5b / fal

Of all the Wan nodes in this pack, this is the one with the least excuse to exist - and it's still useful. The v2.2 5B model was the community's low-VRAM darling: it "fits well on 8GB vram with ComfyUI native offloading," per the release-day thread. So this node runs a model on Civitai's farm (via the fal engine) that a modest local GPU could run itself. The honest use case is the same as ever: no GPU, no VRAM headroom, or you're on a machine where installing a two-pass Wan pipeline is more trouble than a cloud call.

The 5B variant is also the practical pick for this pack: it's cheaper per job than the 14B, and for still images the quality gap is smaller than the size difference suggests. If you're here to make stills, not video, the 5B is the value move.

Inputs and outputs

• prompt (required) - the only required input.
• steps - 40 default, notably higher than the 14B's 27. The 5B dense model compensates for fewer parameters with more sampling steps; that's the two-pass architecture's quirk, and it's baked into the default.
• shift - 2 default, and this one's worth knowing: the sigma shift that balances the high-noise/low-noise passes in Wan 2.2. Leave it alone until you've seen a few outputs; it's the main quality lever and easy to overshoot.
• guidance_scale - 3.5.
• image_size - preset string (square_hd).
• enable_prompt_expansion - off by default; turn it on to let the model flesh out sparse prompts.
• loras, negative_prompt, seed, quantity (max 12), output_format, image_metadata.

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 (CIVITAI_API_TOKEN env var for headless, or the sidebar), and jobs bill in Buzz per image.

Gotchas

Two things trip people up. First, 40 steps default means slow jobs - a 5B model at 40 steps through a cloud queue is not snappy, and the pack's early feedback ("generation times are super slow") applies here harder than anywhere. You can knock steps down for drafts; that's the main tuning lever. Second, don't confuse this node with the v2.2 14B node (CivitaiImageGenWanV22Fal) - different default steps, different cost per job, and the 5B is the one to try first.

And the standing caveat from the Wan story: 2.2 is the last open Wan. Alibaba closed the numbered series after it - 2.5, 2.6, 2.7 are all API-only, which is why the sibling nodes in this pack are cloud-only by necessity. This node is the open-model rung, and if you ever plan to take the same prompt local, the weights are sitting on HuggingFace waiting for you.