Wan 3.0 Pro · Reference to Video

Reference-to-video stopped being a mood-board trick when the models got good enough to hold an identity across cuts. Bernini's multi-reference mode is the open-weights proof - "finally have near closed source level Reference to Video" was the reaction when it landed. This node is the closed-source side of that same capability, at Wan 3.0 Pro quality.

What makes it unusual is the input vocabulary. Most R2V nodes take images and clips. This one also takes reference_file_url - a PDF, DOCX, TXT or MD it reads for characters, wardrobe, setting and action - and reference_link_url for a public page. So the written brief and the visual reference travel in the same request, and the prompt becomes direction rather than description.

Price is Pro-tier: from 21.6 credits ($0.216) per second, 1080p to 4K, 2 to 30 seconds. And it's cloud-only, like everything after Wan 2.2.

How it works

The node is generated from the model's published schema. The two document fields are ordinary STRING inputs - you paste URLs, not file paths, so the target has to be publicly reachable and, for links, readable without a login. The three media fields are autogrow sockets: reference_image_urls, reference_video_urls (up to 5 clips, 1 to 15 seconds each, 15 seconds total) and reference_audio_urls with the same budget. Connect one and another slot appears.

Everything is uploaded when you queue, the run is quoted before credits are held, the task is submitted with an idempotency key, and the node polls with progress text until the MP4 lands.

The inputs that matter

prompt - subject, motion and camera per the tooltip. Short is fine; the references are doing the describing.

reference_file_url / reference_link_url - one each. The generated node's description links the model page, which is where the real constraints on document format and size live.

resolution - 1080p, 2k or 4k, and the tooltip is refreshingly blunt: a native Wan 3.0 render tops out at 1080p, so the 2k and 4k tiers are enlarged from that render rather than generated at those sizes. Buy accordingly.

aspect_ratio defaults to adaptive, which follows your input media and the intent. With a mixed reference set that's usually the right call; if you need 9:16 for certain, choose it.

duration_seconds - 2 to 30, or -1 for smart duration. With reference videos in play, the tooltip notes that their total length plus your output adds up, so a full 15-second reference set and a 20-second ask is a heavy request and a bigger quote.

generate_audio defaults to on. enable_prompt_expansion defaults to on too on the Pro tier - which means your carefully-worded brief gets rewritten unless you turn it off. There's no reasoning toggle exposed on this node either.

Outputs: video and task id.

Installing it

Manager → search PopNow → Install → restart. Or:

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

Add a key from popnow.ai/api/console/keys in Settings > PopNow (or POPNOW_API_KEY), restart again, and the node list rebuilds from your account's catalog. No checkpoints, no dependencies past ComfyUI, no GPU. If custom_nodes/popnow_api.py is still hanging around from the single-file days, delete it.

Where people get burned

Cost compounds quietly here: Pro per-second rates, a 30-second ceiling, and references that each push the quote up. This is the node to put a hard ceiling on - Settings > PopNow > Max credits per run makes the node refuse anything quoted above it before a credit is held.

The other one is the document path. If the URL isn't publicly fetchable, the request fails with a 400 naming the field, and it's easy to misread as a bad prompt. Test your link in a private browser window before you blame the node.

Codes worth memorising across this pack: 40201 no credits, 40004 unsupported combination (names the field), 400 rejected value, 40310 unverified email, content_rejected declined and refunded. Interrupting stops the wait, not the job - the task finishes on PopNow and is charged if it succeeds.