MiniMax H3 Director Group (Image to Video)

The Director node is great at one thing: a timeline you edit inside its own UI. But the ComfyUI way of working is graphs, and sometimes you want each shot of your scene to be a separate node - so you can feed one shot's prompt through a text-rewrite node, pull a shot's keyframe from a file on disk, or reuse the same beat across a batch. That's the job of the Director Group nodes: they let you pack one scene segment as a graph object and hand it to the Director as a batch.

MiniMax H3 Director Group (Image to Video) is the packer for the t2v/i2v/fl2v family. It doesn't sample anything - no model, no VAE, no GPU work. It just bundles a prompt, a duration, and optional keyframes into a typed group that the Director (or a Groups Combine) knows how to swallow.

The inputs that matter

Just two required, which is the whole appeal:

• prompt - that segment's text. It's a normal string input, so you can wire a prompt expander or rewrite node straight into it instead of typing in the Director's UI.
• duration_sec - clip length in seconds (0.2 to 120). The Director snaps it to MiniMax's frame grid at your fps, so a "5 seconds" here means 124 frames at 24fps, the same as the defaults everywhere else in the pack.

Then the two optional image sockets, and this is where the subtlety lives:

• first_frame - connect a first keyframe and the group becomes i2v.
• last_frame - connect a last keyframe and you get fl2v.

What makes this elegant is the combinatorics. No frames wired = the group runs as t2v (pure text-to-video). First frame only = i2v. Last frame alone, or first+last together = fl2v, the first-last frame mode that pins the beginning and end of your shot and lets H3 invent the motion between. Official H3 supports last-only, and this node does too. So you get all three task types out of one node, decided purely by which sockets you fill.

Why you'd reach for it

Three real cases. You're building a workflow where each segment's prompt comes from somewhere else in the graph - a note, an LLM node, a text file - so editing it as a standalone node beats digging through one giant Director card. You want to drag a specific image in as a keyframe without first loading it into the Director's UI. Or you're scripting a batch where the same few segments repeat with different images, and nodes are easier to loop than UI state.

One thing it won't decide for you: output resolution. The Director controls the canvas - width/height live there, not here. A group is just the what of a shot; the Director is the how big.

Wiring and gotchas

A single group goes straight into the Director's i2v_groups socket. Multiple groups first pass through MiniMax H3 Director Groups Combine, which fans them into one list - connect each Group's group output to a Combine slot, then the Combine's groups output into i2v_groups. When external groups are connected, they take priority over whatever's in the Director's UI cards (the cards dim, but "select run" still works to choose which group indices execute).

Do not wire an Image-to-Video group into r2v_groups, and don't mix i2v-family and r2v-family groups in one Combine - the Combine will throw. Match the Director's task_type to the family you're feeding: t2v/i2v/fl2v ↔ i2v_groups.

Otherwise the gotchas are the pack's, not this node's: ComfyUI 0.30.0+ with the official H3 nodes, a minimax-type CLIP, and the right UNET for what you're actually generating (fl2va for this family). This node is the easy, cheap part - the expensive part is what the Director does with your group.