Wan 3.0 Prime · Image to Video

Here's the awkward part of buying hosted video: the node's inputs don't tell you what you're paying for. Prime's schema is identical to the standard Wan 3.0 image-to-video node - same frame input, same 2-to-30-second range, same smart-duration -1, same last-frame pin, same audio - and the price moves from $0.045 to $0.0612 per second. About 36% more for a tier the interface can't describe.

So treat it the way you'd treat any premium tier: run the same frame and prompt through both once, look at them side by side at 100%, and decide with your eyes. If you can't see the difference on your material, save the third. If Prime is what makes hands and fast motion hold together, pay it for the takes you keep and draft on the cheaper node.

What the inputs do

image_url is the opening frame - everything about the look is decided there. prompt carries motion: action, camera, pacing. The I2V rule (describe what happens after the still, don't re-describe the still) is worth more than any parameter on this node.

duration_seconds defaults to 5, and -1 means smart duration - the model chooses the length from the prompt and media instead of a fixed number. That's the reason to prefer the 3.0 family over 2.6 in the first place. resolution is 480p, 720p or 1080p; 480p here is genuinely useful for testing a motion idea cheaply, which the 2.6 nodes don't offer.

aspect_ratio defaults to model default: omit it and the task adapts to your opening image instead of imposing a shape. Set it only when the output has to fit a format.

generate_audio is on by default - the clip comes back with sound. enable_thinking lets the model reason about the brief and reference material first; it's slower, better with complicated references, and switches itself on when a reference file or link is present.

enable_prompt_expansion is off by default. seed repeats a draw exactly. The optional last_image_url pins the final frame and requires image_url to be connected - it's how you make two clips meet on a deliberate cut rather than an accident.

Outputs: video → Save Video or into an extend/upscale chain; task id as a string for the console and support.

Install

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

Restart, then Settings > SpicyAPI with a key from spicyapi.ai/console/keys - or search "SpicyAPI" in ComfyUI Manager. No model files, no dependencies; the HTTP client is aiohttp from ComfyUI's own environment, so there's nothing here that can break a working install. No GPU and no VRAM either: the node uploads, waits, and downloads. Your key sits in the ComfyUI user folder rather than in a widget, which keeps it out of workflow files and image metadata; use SPICY_API_KEY on a headless server.

Because Prime costs more, be deliberate

Set Max cost per run in the SpicyAPI settings. The node badge is a per-second rate, and at this tier a 30-second 1080p run is real money - the ceiling turns an accidental long duration into a stopped request instead of a charge.

Draft on the standard 3.0 node or at 480p, then spend Prime's rate on the final take. Keep the seed control on fixed so re-queueing reuses a result rather than buying a second one, and change it to randomize only when you want a new draw. Since a cancelled run continues server-side and is charged if it succeeds, cancelling is not a cost control - your results stay in the console for 14 days, so check there first.

Errors worth recognising: 40004 for an option combination with no deployment (the message names the parameter), 40003 when an upload didn't match its ticket (re-run the node), 40310 for an unverified email, 50301 when no deployment is available right now, 50302 when the generation failed upstream and was refunded and can just be re-run, and No SpicyAPI key is set when the key is missing.