MiniMax H3 Director Group (Reference to Video)

H3's reference mode is how you keep a subject locked across shots - feed it reference images of the character or object you want, and each clip it generates is supposed to stay that same thing. Inside the Director, that whole "reference pack" is built in the node's own card UI. MiniMax H3 Director Group (Reference to Video) is the graph-native version: one node per reference shot, which you can then batch, combine, and feed to the Director alongside all the other tools in your workflow.

Like its i2v sibling, this is a pure packer. No model, no VAE, no sampling. It bundles a prompt, a duration, and any number of reference media into a typed group for the Director's r2v_groups socket.

The inputs that matter

prompt and duration_sec are the only required inputs - prompt, plus clip length snapped to H3's frame grid (5 seconds ≈ 124 frames at 24fps). The prompt is where H3's tag syntax comes in: you reference your loaded media inline as <Picture N>, <Video K>, and <Audio J>. This is the sentence to remember: the tags only mean something if the matching slot is actually filled.

Then four Autogrow inputs, which is what makes this node feel like the official H3 reference node:

• ref_images - reference images, up to 9, becoming <Picture 1> through <Picture 9>.
• ref_videos - reference video frame batches, up to 3, becoming <Video 1>+.
• ref_video_audios - a soundtrack paired with the same-numbered reference video.
• ref_audios - standalone reference audio, up to 3, becoming <Audio J>.

Autogrow means you plug an image in and a new empty slot appears, exactly like the official MiniMax H3 Reference to Video node - no fixed nine-ports wall to scroll past. And note that this is a reference to video group, so the picture stays character/subject-consistent per shot, which is what makes a multi-shot edit feel like one production instead of a montage of strangers.

Wiring it up

Each Group node emits one group. A single group can go straight into the Director's r2v_groups socket; for a multi-shot scene, chain several Groups into MiniMax H3 Director Groups Combine first, then feed the Combine's groups list to the Director. Once external groups are connected they override the Director's UI cards for execution - same external-priority rule as the i2v side.

Two constraints to keep straight. Don't mix Reference-to-Video groups with Image-to-Video groups in one Combine (it errors). And connect to the matching socket: r2v groups belong on r2v_groups, and the Director's task_type should be r2v (or v2v/rv2v, which use the same reference machinery).

What it assumes you have

This is the heavier side of the H3 pack, so double-check the prerequisites. Reference mode routes through the official MiniMaxH3ReferenceToVideo node, which needs ComfyUI 0.30.0+ and - this is the one people miss - the audio_vae connected to the Director, because reference video with audio is part of the deal. The Director's UNET should be the ref2va variant, not fl2va, for the reference path. And the standard H3 housekeeping still applies: minimax-type CLIP for text encoding, and the territory-license check if you're in the US, EU, UK, or South Korea.

The node itself is the easy part of that chain. If a shot comes back wrong, the first place to look isn't the packer - it's whether the reference actually matches your <Picture N> tag, and whether the right UNET is in front of the Director.