Wan 3.0 Pro Prime · Text To Video

Text in, video out, audio included, no source image anywhere in the graph. That's the whole node. Wan 3.0 Pro Prime sits at the top of Alibaba's hosted 3.0 family, and SpicyAPI exposes its text-to-video endpoint with the same knobs the API takes: a prompt, a resolution tier, a ratio, a duration, and a couple of switches.

When text-to-video is the right call

If you have a still, image-to-video is almost always the better spend - the composition is decided, and you're paying the model to do motion instead of doing everything. Text-to-video earns its rate on shots you can't photograph: camera-only moves, weather, crowds, a location that doesn't exist. It's also the fastest way to test whether a hosted video model is worth your money at all, because four prompts cost less than four hours of tuning.

The honest caveat that applies to every node in this pack: generation happens on SpicyAPI's servers, so your prompts and reference material leave your machine and the model's own policy is the one enforcing anything. That's the price of a model you can't download. For the open Wan 2.x family, you have a local option and should probably use it; for 3.0, you don't.

The controls

• prompt - subject, motion, camera. Pro Prime defaults prompt expansion on, which means the service rewrites your brief before rendering. Nice for lazy prompts, annoying when you're testing a specific phrasing; flip it off to hear yourself think.
• resolution - 1080p (the default), 2k or 4k. The author's tooltip is candid: native Wan 3.0 Prime renders stop at 1080p, and 2K/4K are enlarged from it. If you need real detail, render at 1080p and upscale locally, where you choose the method.
• aspect_ratio - 16:9, 4:3, 1:1, 3:4, 9:16 or adaptive (the default). adaptive lets the intent of the prompt decide the frame, which is usually better than guessing.
• duration_seconds - 5 seconds by default, or -1 for smart duration: the model picks a length from the prompt instead of cramming a 20-second idea into five seconds.
• generate_audio - on, and it's the reason many people are here. The sound arrives inside the same file.
• seed.

Outputs: video → Save Video, and task id → a STRING you can quote to support when a run dies.

What it costs

$0.234/s at 1080p, $0.252/s at 2k, $0.279/s at 4k. Five seconds of 1080p is about $1.17; twenty seconds is $4.68; a 30-second 4K shot is $8.37. That's not a rounding error, so treat this node the way you'd treat a render farm: draft cheap, commit expensive. Draft in the base Wan 3.0 t2v node at $0.045/s, then re-run the winning prompt here.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Or search SpicyAPI in ComfyUI Manager. Restart, then Settings > SpicyAPI → paste a key from spicyapi.ai/console/keys and restart once more, because the node list is rebuilt from your account's catalogue at startup. Headless server? Use SPICY_API_KEY. There's nothing to install besides the node pack - no model weights, no extra Python packages (Python 3.10+ and the V3 node API, tested with ComfyUI 0.39).

Guardrails

Set Max cost per run in the SpicyAPI settings before you queue anything expensive. The pack quotes the request and refuses to submit if the worst case breaks your ceiling - a real quote, not an estimate from the widget values, which matters at this tier. Leave the seed on fixed while you iterate so a re-queue reuses the first result instead of silently buying a second one, and remember that cancelling a run doesn't cancel the job: it finishes on the server, is charged if it succeeds, and stays in your console for 14 days.

If the node won't run, the message will tell you why: No SpicyAPI key is set, 40310 (verify your email), 40004 (no deployment serves that combination - change the option the message names), or an error tail carrying a request_id, which is what support needs to find the request.