Wan 3.0 · Reference to Video

Most "reference to video" nodes mean "one photo of a person, do something with it". This one means something stranger and more useful: point it at a document or a web page and it reads that for characters, wardrobe, setting and action before it renders. Alibaba calls that a reference, and Wan 3.0 is the tier that does it without the Pro price.

If you've ever written a four-paragraph prompt describing a scene you already have written down somewhere, this is the shortcut.

The inputs, which are unusual

Start with the two text references, because they're the point.

reference_file_url is a document upload: PDF, DOCX, TXT, MD and similar. In the node it's a file field, so put the document in ComfyUI/input and type the filename, or paste an https:// link. That file is uploaded and read as reference material. reference_link_url is a public web page - it has to open without a login, or the model gets a paywall instead of your setting notes.

Then prompt. The tooltip asks for subject, action, camera, light, pacing and boundaries, which is a good checklist for a shot you intend to nail in one go.

enable_thinking is the switch that makes this node feel different. Turned on, the model reasons about the brief and the reference material before it generates - slower, and noticeably better on complicated references. It also switches itself on when you supply a reference file or link, so you mostly leave it alone.

resolution here is 480p, 720p or 1080p, and aspect_ratio is 16:9, 9:16, 1:1, 4:3, 3:4 or adaptive. generate_audio, as everywhere in this pack, is on by default. enable_prompt_expansion has a third state - it can be left at "model default" and omission means the service decides.

Media references are sockets, not files

reference_image_urls, reference_video_urls and reference_audio_urls are autogrow inputs: connect an IMAGE, VIDEO or AUDIO and the socket grows a sibling for the next one. Limits are in the tooltips - up to 10 images, up to 5 videos and 15 seconds total, up to 5 audio clips and 15 seconds total.

That 15-second video cap is not decoration. If your reference videos total 20 seconds, the API refuses the job; and because the reference video seconds are billed alongside the output seconds, and the two together must fit under 30, a fat reference reel eats your output length from both ends.

How it runs

Same engine as every node in this pack: widgets mirror the published schema of alibaba/wan-3.0/reference-to-video, media gets uploaded to popnow:// file URIs that only your key can use, the run is quoted before anything is held, submitted under an idempotency key, polled, then downloaded and decoded into a VIDEO. The second output is task id - your handle for the API console and for support.

Cost starts at 6.75 credits per second, 1 credit = $0.001, so five seconds lands near 34 credits. Reference material adds billed seconds when it's video.

Install

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

Or search PopNow in ComfyUI Manager. No dependencies, no model files, no GPU. Then create a key at popnow.ai/api/console/keys, restart, paste it under Settings > PopNow or export POPNOW_API_KEY, and restart once more so the node list comes from your account.

Where people get burned

The required text fields are genuinely required. Leave reference_link_url empty and the node stops with an "is required" error before a request is ever made - even when your reference material is a stack of connected images. Put something real in it, or accept that this node wants a link.

The other trap is expecting a document reference to behave like a style LoRA. It doesn't copy a look; it reads for content. Give it a script and you get the scene in the script. Give it a mood board and you'll wonder why the output is generic.

One last thing: cancellation ends the wait, not the job. A cancelled run finishes server-side and is charged if it succeeds. Check the API console before assuming a cancel was free.