Wan 3.0 Prime · Reference to Video

Every other node in this pack takes media. This one takes documents. There are two plain string fields on it - reference_file_url for a PDF, DOCX, TXT or MD, and reference_link_url for a public web page - and the model reads them for characters, wardrobe, setting and action before it renders.

That is a genuinely different workflow. Instead of writing a 400-word prompt describing who's in the shot and what they're wearing, you write the brief once as a document, point the node at it, and supply the visual references separately: up to 10 stills, up to 5 clips totalling 15 seconds, up to 5 audio clips totalling 15 seconds. Character sheets, mood boards, voice samples - all in one request. It's the closest thing to "here's the production bible, make me a shot" that I've seen in a ComfyUI node.

The tier is Wan 3.0 Prime, so: 480p/720p/1080p, 2 to 30 seconds, native audio, 9.18 credits ($0.0918) per second, rendered quicker than base 3.0. And no weights, per the whole post-2.2 story.

How it works

The pack builds this node from the endpoint's schema, so the document and link fields arrive as ordinary STRING inputs - you paste a URL, not a file dialog, which means the doc has to be reachable and the page has to open without a login. The three media fields are autogrow sockets: connect one image and a second slot appears, and so on. Everything you wire in is uploaded for you.

Then the usual: quote before credits are held, task created under an idempotency key, progress printed on the node, MP4 downloaded as a VIDEO.

The inputs that matter

prompt is still required, and it's the direction rather than the description. Use it to say what this shot is - "the confrontation in the warehouse, handheld, cold light" - and let the document supply the specifics.

reference_file_url and reference_link_url take one each. The tooltip for the link is explicit that it must open without a login, which rules out most useful internal documentation and all of the good stuff behind Notion auth.

enable_thinking is worth turning on here. It's off by default, but a document plus a handful of references is precisely the case the tooltip describes as benefiting from a reasoning pass - and there's a nice detail in the schema: thinking switches on by itself when a reference file or link is given. You can still force it either way.

enable_prompt_expansion is a dropdown on this node rather than a switch, defaulting to model default - i.e. the field isn't sent at all and the model decides. Don't go looking for a boolean that isn't there.

duration_seconds - 2 to 30 or -1 for smart. With reference videos in play, their total length plus your requested output adds up, so a 15-second reference set and a 20-second ask is a heavy request. resolution - 480p, 720p, 1080p. aspect_ratio - five fixed ratios plus adaptive.

Outputs: video and task id.

Installing it

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

or search PopNow in ComfyUI Manager and install. Restart, add a key from popnow.ai/api/console/keys under Settings > PopNow (or POPNOW_API_KEY), restart once more so the catalog loads. No models, no dependencies, no GPU.

Remove the legacy custom_nodes/popnow_api.py if it's still installed.

Where people get burned

Dead links and unreadable documents. A 404 or a login wall on reference_link_url fails the run, and because the failure comes back as a 400 naming the field rather than a friendly "couldn't read that page", people assume the node is broken. Host the PDF somewhere public and paste the direct URL.

Reference budget creep is the other one. Ten images, five clips and five audio files is a lot of upload for one request, and every reference raises the quote. Start with three.

Interrupting the wait instead of stopping the spending is the classic mistake across this whole pack: the task keeps running server-side and settles if it succeeds. Set Settings > PopNow > Max credits per run to a number you'd shrug at. Credits errors are 40201, unsupported combinations are 40004 with the field named, refusals are content_rejected and refunded.