RunPod WAN 2.2 T2V 720p

Wan 2.2 is the quality standard of open video - the MoE model with more expressive motion, better emotion, and sharper detail than 2.1. It's also a 27B parameter monster that will make your GPU sweat and your RAM cry. This node runs it as a hosted text-to-video call, so you get the 2.2 generation without the 27B download or the 50–80GB memory profile that local 2.2 often demands.

Let's be clear about what you're buying, though. Wan 2.2's weights are open - this isn't a locked model like Veo or Sora. This is the "I don't have the hardware / don't want the hassle" path (per-call billing, prompts leaving your machine, a server-side safety filter you can't edit). If your only barrier is VRAM, this node is the cheapest way across it short of renting a pod. If you already run 2.2 locally, you probably don't need it - but for a quick prompt-to-clip test on a laptop, it's hard to beat.

How it works

Standard pack mechanics: the node wraps RunPod's wan-2-2-t2v-720 endpoint, submits an async job with your API key in the header, and polls the status URL once a second until COMPLETED. The ComfyUI console streams [RunPod] Status: ... lines while you wait - 30 seconds to a couple of minutes depending on load. On RunPod's side, the endpoint runs the two-pass MoE pipeline: the high-noise expert handles motion and composition, the low-noise expert refines detail. You don't see any of that; it's the server's business.

Because Wan is flow-matching, the knobs are guidance and flow_shift (defaults 5 and 5, the right neighborhood) rather than CFG. If your image-model habits scream "crank CFG," don't.

The inputs that matter

• api_key - your RunPod key, typed into the field (a Primitive node lets you share one across a graph).
• prompt - the entire scene in prose. The default - "A serene morning in an ancient forest, golden sunlight filtering through tall pine trees" - is a solid template: subject, setting, light, mood.
• duration - 1–10 seconds, default 5.
• seed - -1 random, or lock it for a repeatable take.

The rest (negative_prompt, size 1280*720, num_inference_steps 30, guidance, flow_shift, enable_prompt_optimization, enable_safety_checker) have sane defaults. One honest note: T2V is the harder mode. Without an anchor image, consistency and motion are less predictable than I2V - if you have a starting frame, the I2V node will usually serve you better.

The output

One output: video_url, a STRING. It connects to a "Load Video from URL" node or a Save Text node, not to a video player - and the URL expires after about seven days. Download anything you want to keep, and treat any stored link as temporary.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/tcarwash/ComfyUI_RunpodNodes

Restart, or install "RunpodNodes" from ComfyUI Manager. No requirements.txt, no heavy dependencies, no models to fetch - the pack only uses requests, PIL, torch, and numpy, all already present. That's the pack's best feature: the whole install is one clone.

The usual API-node caveats apply: paid per call (video runs roughly $0.30–1.20 per the README), your prompt goes to a third party, and this pack is a new single-author release with no track record yet - skim the source before you trust it with a key. For "I want a 2.2 clip and I don't want to fight my GPU for it," this is the node.