Wan 3.0 Pro Prime · Reference To Video

Character consistency is the oldest unsolved problem in this hobby. IP-Adapter, FaceID, PuLID, face LoRAs - all of that machinery exists to get the same person into the next image. Reference-to-video is the hosted answer: hand the model a handful of images, a couple of clips and some audio, and it keeps the identity, wardrobe, set, movement or voice while generating a new 2-to-30 second shot. This is the Pro Prime tier of that, at $0.234/s and up.

What makes it different from image-to-video

Image-to-video animates a frame you already made. Reference-to-video takes material about a subject and generates a scene the subject was never in - closer to what the KB calls the 2026 answer to consistency, where the model looks at the reference while it works instead of re-inventing the character from a prompt. It's the right tool for a series: a product, a presenter, a recurring character, a location that has to read as the same location three clips in a row.

The inputs

Three required reference fields sit at the top, which is unusual and worth knowing before you open the node:

• reference_file_url - one document (PDF, DOCX, TXT, MD and similar) that the model reads for characters, wardrobe, setting and action. Give it a plain filename that lives in ComfyUI's input folder, or a public https link; the pack resolves bare names inside that folder only.
• reference_link_url - one public web page, and it has to open without a login.
• prompt - subject, motion and camera, as usual.

Then the usual shot controls: resolution (1080p default, 2k, 4k - enlarged from a 1080p native render, so paying 4K rates for an upscale is the least sensible thing to do here), aspect_ratio (adaptive by default), duration_seconds (default 5, -1 for smart duration), generate_audio, enable_prompt_expansion (on by default), seed.

The interesting part is the reference media. reference_image_urls, reference_video_urls and reference_audio_urls are autogrow sockets: connect one image and a new socket appears, up to 10 images, 5 videos and 5 audio clips. Caps worth memorising - reference videos are 1 to 15 seconds each and 15 seconds in total, same total for audio.

Outputs: video for Save Video, and task id for the console or a support ticket.

The billing trap

Reference videos are billed together with your output seconds. Feed it 15 seconds of reference footage and ask for a 15-second render and you're paying for 30 seconds at the Pro Prime rate - $7.02 at 1080p, before the 2K/4K uplift. The schema spells this out on the duration tooltip; the reference length plus the output length must stay under 30 seconds anyway, so plan the shot around that ceiling rather than discovering it in the receipt. Images and audio don't bill the same way, so a stills-and-voice reference set is much cheaper than a video one.

Install

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

Manager works too - search SpicyAPI. Restart, then Settings > SpicyAPI, paste a key from spicyapi.ai/console/keys and restart once more so the node list comes from your account. SPICY_API_KEY for servers. Nothing to install and no weights to download: the pack's dependency list is empty and it only uses what ComfyUI already ships (Python 3.10+, V3 node API, tested on 0.39).

Housekeeping

Set Max cost per run before your first reference run - reference billing is the least predictable thing in the pack, and the pre-flight quote is what stops a fat-fingered 30-second 4K request. Seeds start fixed, so a re-queue is free (same result, cached); randomize buys a new take. Cancelling only stops your wait - the job finishes server-side and is charged if it succeeds. Errors: No SpicyAPI key is set, 40310 unverified email, 40004 no deployment for that combination, and any message ending in request_id ... is worth sending to support verbatim.