Modelverse Wan-AI T2V

If you follow local video generation, you know the Wan story: Alibaba's Wan 2.1 and 2.2 are the open models that put video on every GPU owner's machine - a 14B diffusion transformer, a 3D causal VAE, MoE - and then Alibaba froze the open line and pushed the numbered series behind an API. Modelverse Wan-AI T2V is that API, wrapped as a ComfyUI node: the full Wan 2.2 text-to-video model, hosted, with none of the multi-gigabyte downloads or the VRAM math.

That's the pitch, and it's a good one. Local Wan 2.2 needs a serious card and an afternoon of setup. Here you type a prompt, pick a resolution, and the model runs on UCloud's servers while your GPU idles. The KB's Wan panel calls 2.2 the "last open Wan" - so if you want the closest thing to that model without the hardware commitment, this node is the shortcut.

The inputs that matter

• client - from the Modelverse Client or Secret Client node. Standard pack wiring.
• prompt - the scene you want. This is the whole job; describe motion, camera, subject.
• negative_prompt - defaults to low quality. Wan actually honors this, so it's worth using ("low quality, blurry, distorted").
• resolution - 720P (default) or 480P. 720 is the sweet spot for cost and quality; 480 is for cheap drafts.
• size - the frame geometry: 1280x720, 720x1280 (portrait), 832x480, 480x832. Pick the one matching where the video will live.
• seed - 0 to 2147483647 for reproducibility.

One honest limitation from the source: duration is fixed at 5 seconds. There's no duration input on this node; it's hardcoded into the request. If you need longer clips, that's what the Vidu Extend or a different model is for.

How it works

The node submits a task to Modelverse's /v1/tasks/submit endpoint as Wan-AI/Wan2.2-T2V, then polls /v1/tasks/status every 5 seconds until the video is done. That's the pattern all the video nodes in this pack share, and it means a run blocks while the video renders - 5-second Wan clips typically take minutes. The console shows the task status ticking along so you're not staring at a frozen UI.

The outputs are the two strings you'll wire everywhere:

• url - the finished video's URL. This feeds Modelverse Preview Video to save and preview it as a proper VIDEO object.
• task_id - the task identifier, handy for debugging or for polling the Modelverse console.

Install and expectations

Standard pack install: ComfyUI Manager → search "ComfyUl-UCloud" → Install → restart, or git clone https://github.com/ucloud/comfyui-plugin into custom_nodes. Dependencies are light; the "heavy" part is purely the cloud bill.

Be realistic about it: this is a metered, hosted 14B model, so generations aren't free and queue times vary. But if the alternative is "I don't own a 24 GB card, so I can't run Wan at all," this node makes Wan 2.2 the easiest part of your day.