wan / v2.2 / comfy

This is the node for people who feel naked without a sampler dropdown. It runs Wan 2.2 - the quality-first release that took over local video - on Civitai's cloud through a ComfyUI-style engine, and it hands you the knobs a local KSampler has: 18 samplers, 7 schedulers, steps, CFG, and a shift parameter. Prompt in, native VIDEO out, billed in Buzz. It's the most "this is just ComfyUI, but somebody else's GPU" node in the official civitai-comfy-nodes pack.

If you've run Wan 2.2 locally, this node needs almost no translation - it's the same mental model with the local loaders swapped for Civitai's. If you haven't, the practical truth is you can leave nearly every knob alone and still get good clips; the "comfy" engine exists so power users feel at home, not because defaults are broken.

The inputs that matter

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

• sampler - 18 choices, default euler. The full ComfyUI roster: euler, euler_ancestral, heun, heunpp2, dpm_2, dpm_2_ancestral, lms, dpm_fast, dpm_adaptive, dpmpp_2s_ancestral, dpmpp_sde, dpmpp_2m, dpmpp_2m_sde, dpmpp_3m_sde, ddpm, lcm, uni_pc, uni_pc_bh2. euler is the Wan community's default for good reason; don't migrate to lcm or uni_pc unless you know why.
• scheduler - normal, karras, exponential, sgm_uniform, simple, ddim_uniform, beta, default simple.
• shift - 1.0 to 20.0, default 8.0. This is the Wan-specific training-shift parameter; it controls where the model spends its denoising effort. Leave it at 8 unless you're tuning deliberately.
• width / height - default 1280×720, min 64, max 2048. Set the exact frame you want.
• steps - 10 to 50, default 20.
• model - a CIVITAI_AIR socket from the Civitai Model Selector, so you can run a specific Wan 2.2 checkpoint instead of the platform default.
• loras - CIVITAI_LORAS socket; trigger words go in the prompt (the selector's keywords field is just a reminder).
• source_image / images - image-to-video inputs.
• negative_prompt - worth using here, since you have full CFG control.
• duration 1–30s (default 5), frame_rate 24, seed, api_config.

Outputs

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

Install and auth

One-time pack 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 (headless), or sidebar sign-in.

Troubleshooting and takes

• The model socket won't fill. It's not a text widget - wire the Civitai Model Selector's air output into it. Same pattern as every CIVITAI_AIR/CIVITAI_LORAS socket in the pack.
• Sampler rabbit holes. With a cloud meter running, every experiment costs money. Wan 2.2 is a two-pass architecture and its sampler behavior doesn't mirror SD 1.5 habits; stick with euler/simple and tune steps and shift before blaming the sampler.
• High resolution = long queue. 2048px is legal here, but this pack launched June 2026 during Civitai's GPU crunch - big frames plus busy queues is the timeout recipe. Default timeout 30 min; raise via Civitai Auth timeout_minutes or CIVITAI_COMFY_TIMEOUT.
• Negative prompt does nothing sometimes. If your negatives seem ignored, remember the model's content filter runs server-side regardless - that's Civitai's layer, not the CFG.
• Early preview. Generated nodes "may change without notice"; re-check the sampler list after updates.

Reach for this node when you want Wan 2.2 quality and the feel of driving the sampler yourself. If you'd rather not think about samplers at all, the fal-hosted Wan 2.2 nodes give you the same model with far fewer dials.