Wan 2.6 · Reference to Video

Keeping a face stable between two shots is the problem everyone in this hobby hits and nobody has cleanly solved. The local answer in 2026 is instruction editing - generate the character once, then edit that image into new poses - and for stills it works. For motion it falls apart, because now you need the identity to survive and the movement to be coherent across a cut.

That's the gap this node fills. Wan 2.6 reference-to-video takes reference clips and stills of your subjects and generates a new scene containing them. Not an animation of your frame; a fresh shot with your cast in it, optionally cutting between setups.

The catch is the same as the rest of the Wan 2.x line past 2.2: closed weights, API only. If you were hoping to run 2.6 at home, you can't - Alibaba stopped publishing checkpoints after 2.2 and the community has been sour about it since 2.6 shipped in December 2025. You queue, you pay credits, you get a clip back.

How it works

One node per model, built from the model's own input schema. The two multi-file inputs - reference_video_urls and reference_image_urls - are autogrow sockets: connect one clip and a second socket appears, and so on up to the cap. Everything you wire in is uploaded for you and appears to the model as an ordered list.

Order matters, and the author says so in the prompt tooltip: describe the new scene, naming each reference and saying what it contributes, then the setting, action, camera and light. Video references are numbered first, stills after them. Up to 3 reference clips, up to 2 stills.

Cost is per second of output, from 9.75 credits ($0.0975) - cheaper than the 13.5 base tier, which is unusual and probably tells you something about how the endpoint is capacity-priced.

The inputs that matter

prompt plus your references is the whole request. There's no required image socket at all.

shot_type is the one I'd point a beginner at first: single gives one continuous take, multi lets the model cut inside the clip, and multi-shot is where reference-to-video earns its keep - cut from a wide to a close-up and the same person is still there.

duration_seconds is a straight integer, 2 to 10. resolution gives 720p or 1080p. aspect_ratio defaults to 16:9 and is set independently of your references - pass a 9:16 clip as a reference and the output is still 16:9 unless you say otherwise.

negative_prompt and seed behave as you'd expect. Note there's no prompt-expansion control on this variant; 2.6 reference-to-video doesn't take one, so what you type is what gets sent.

Outputs: video into Save Video, and task id for support or the API console.

Installing it

Search PopNow in Manager's custom node list and install, or:

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

Restart, then add a key from popnow.ai/api/console/keys in Settings > PopNow - or export POPNOW_API_KEY on a headless box - and restart once more so the node list rebuilds from your account's catalog. No models to download, no dependencies beyond ComfyUI. If you previously used the single-file popnow_api.py, delete it; the node types are the same and having both is asking for trouble.

Where people get burned

Treating reference-to-video like image-to-video. Your first reference is not the first frame. If you need the clip to open on exactly that picture, you're on the wrong node - use Wan 2.6 image-to-video.

Under-specifying the prompt is the other common one. Two references and "make a video of these characters" gives mush; "character1 walks into the diner, character2 looks up, slow push in, hard afternoon light" gives you a shot.

Error codes are all in the README's troubleshooting table: 40201 is credits, 40004 is an unsupported combination of settings, 400 names the field it didn't like. Refused or failed runs are refunded. And interrupting the queue won't cancel the job - it keeps running server-side and settles if it succeeds, so your spending ceiling (Settings > PopNow > Max credits per run) is the lever that actually protects you.