Modelverse Wan-AI I2V

Image-to-video is where the interesting video work actually happens. A prompt-to-video roll of the dice might give you a weird scene you never pictured; start from an image you already like and the model's job is narrower - animate this. Modelverse Wan-AI I2V is Alibaba's Wan 2.2 image-to-video model hosted on UCloud's ModelVerse API, which means you get the model the community treats as the reliable open base (per the KB's Wan panel) without downloading a single checkpoint or owning the GPU to run it.

The natural workflow: generate a still with one of the pack's image nodes - Qwen Image, NanoBanana, whatever you have - then drop that still into this node and let Wan 2.2 give it motion. That pairing is the most genuinely useful loop in this entire pack.

The inputs that matter

• client - standard wiring from Modelverse Client or Secret Client.
• prompt - defaults to "Convert to video" but you should write something real. Wan uses it to decide how the scene moves: "slow zoom into the window, leaves drifting" beats "make a video."
• first_frame_image / first_frame_url - one of these is required, and the node refuses both at once. The first frame is the whole point of I2V. Wire an IMAGE tensor into first_frame_image normally; first_frame_url is for when your source already lives on the web.
• last_frame_image / last_frame_url - optional, and genuinely nice: give Wan an ending frame and it animates between your two stills, for a start-to-end transition.
• negative_prompt - defaults to "low quality, blurry"; leave it or extend it.
• resolution - 720P (default) or 480P.
• seed - for reproducibility.

Like its text-to-video sibling, duration is fixed at 5 seconds - it's hardcoded in the request, so don't go looking for a duration dial.

How it works

The node base64-encodes your image tensor (or passes through your URL), submits a task as Wan-AI/Wan2.2-I2V, then polls every 5 seconds until the clip is done - up to about 10 minutes before it gives up. The two outputs are url and task_id, and url is the one you'll wire into Modelverse Preview Video to save and watch the result.

A detail worth knowing from the source: if you provide neither a first-frame image nor a URL, it raises a clear error - and providing both raises a different clear error. The both-or-neither validation is intentional, so when you hit it, it's not a bug, it's the node keeping you from sending garbage to the API.

Install

The pack installs once for all these nodes: ComfyUI Manager → search "ComfyUl-UCloud" → Install → restart, or git clone https://github.com/ucloud/comfyui-plugin into custom_nodes. Light dependencies, no model downloads, and a Modelverse API key from the console.

The usual caveat applies: hosted, metered, and slow-ish - a 5-second I2V clip can take a few minutes of polling. That's the cost of not owning a 14B video model.