MiniMax H3 · Reference to Video

Keeping a character recognisable across shots is the problem this whole hobby keeps tripping over. The 2026 answer has mostly been "generate once, then edit that image" - but that only gets you stills. Reference to Video is the moving-picture version: point it at photographs, a couple of clips and a voice sample, and ask for a new shot featuring those subjects.

What it is

A PopNow cloud node for MiniMax H3's reference-to-video endpoint, starting around 5.55 credits ($0.0555) per second of output. Nothing local, no LoRA training, no IP-Adapter, no face swap. You hand it references, you describe a shot, it renders that shot with those subjects in it. In the PopNow pack, it sits under PopNow > Video > MiniMax next to the plain text-to-video and video-extend H3 nodes.

How the mechanism actually works

The three reference groups are sockets that grow. Connect one image to reference_image_urls and a second socket appears; the pack builds them as ComfyUI Autogrow inputs named reference_image_0, reference_image_1 and so on, up to nine. Same for the video and audio groups. That's why the node looks different in every screenshot - it's shaped by what you plug in.

You then refer to the references by number in the prompt. H3 treats them as conditioning, not as a first frame: the shot doesn't start at your photo, it starts wherever your prompt says, with those subjects cast in it. That's the difference between this node and the image-to-video one, and it's the reason it's the more useful of the two for narrative work.

Files are uploaded for you - images encoded to PNG (JPEG if they'd exceed the 10 MiB ceiling), audio as 16-bit WAV, videos as MP4 - and cached for 12 hours, so iterating on the prompt doesn't re-upload your references each time.

The inputs

• prompt - name the references by number, then describe the shot they should produce. "Image 1 walks into Image 2's kitchen, camera tracks right, kettle starts whistling" is the shape that works.
• reference_image_urls - up to nine. This is where faces, products, costumes and locations go.
• aspect_ratio - 16:9 default; also 9:16, 1:1, 4:3, 3:4, 21:9.
• resolution - 768p or 2k, default 768p.
• duration_seconds - 5 to 15.
• reference_video_urls (optional) - up to three clips, 2 to 15 seconds each, 15 seconds total. Refer to them as <Video 1> and so on. Their seconds are billed at a separate, lower per-second rate than output seconds.
• reference_audio_urls (optional) - up to three WAV or MP3 clips, 2 to 15 seconds each and 15 total, for a voice or a specific sound.

Outputs are video (into Save Video) and task id for the API console.

Casting with audio is the underrated part

Because H3 generates sound natively, a voice reference isn't decoration - it's the difference between "a man speaking" and that man's voice in the shot. Three seconds of clean speech, no music, no room echo, is enough. If you're building a character for a series, keep a folder with one portrait, one full-body shot, one costume detail and a five-second voice sample. That's your casting kit, reusable across every shot you make with this node.

Install

Manager → search PopNow → 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 and restart again. No extra Python packages - the pack's dependency list is empty.

Where it goes wrong

Too many references is the classic mistake. Nine images of the same person in nine different lighting conditions gives the model nine different people to average; three good ones beat nine mediocre ones. Each reference also nudges the quote up, so more isn't free.

PopNow 400 naming a field means a value the model won't take - usually an aspect ratio or duration outside the listed range. PopNow 40201 is credits. And since the prompt is all you have to aim this thing, write it before you attach anything; the references support a shot description, they don't replace one.