VELVET VICE MiniMax H3 — Director

If you open the Velvet Vice MiniMax H3 reference workflow and start poking, the VelvetViceMiniMaxH3Director is the node that feels like the cockpit. It's where you tell H3 what you're making: how long, what shape, what resolution, whether it talks, and what kind of ending it lands on. Everything technical - which model file, which backend, the actual loaders - is deliberately pushed into the System Hub so the Director stays a place where you think about the video, not the plumbing.

The pack's README frames it as "the main MiniMax H3 user director: duration, automatic image format and resolution, seed control, Turbo master switch, audio-output control and ending mode." That's the honest summary. Let's go through the handful you'll actually touch.

The inputs that matter

• first_frame (IMAGE) - your starting image, the I2V seed of the whole clip. Nothing renders without it.
• duration_seconds (FLOAT, default 5) - requested length. Here's the H3 quirk: the model doesn't render an arbitrary frame count. It snaps to its native 17k+5 frame grid at 24 fps (so 3592 frames ≈ 149.67s is the hard cap), and the commonly trained range is roughly 5–15s. Ask for 7 seconds and H3 rounds onto its grid. Anything past ~15s is genuinely untested territory upstream, so don't be surprised by drift if you push 60.
• format_mode - AUTO (read the image's own aspect), 16:9, 9:16, 1:1, or CUSTOM.
• resolution_mode + resolution_preset - preset sizes (900p default, up through 1440p) or custom megapixel / exact-size overrides. H3 does up to 2K, but your VRAM is the real ceiling; the QUALITY preset is where that trade lives.
• render_preset - TEST / FAST / BALANCED / QUALITY / CUSTOM, the author's tier ladder.
• turbo_enabled (BOOLEAN) - the master switch for the Turbo/Distilled LoRA path. Off = base model at ~20 steps; on = fast step count, but only if a compatible Turbo LoRA is actually detected.
• native_audio_output (BOOLEAN, default true) - WITH SOUND vs MUTED · VIDEO ONLY, propagated all the way through VAE loading and decode.
• ending_mode - AUTO / NO CLIMAX / CLIMAX / LOOP / CONTINUOUS ACTION. H3's omni-context lets you steer how the shot finishes, which is one of those features that sounds gimmicky until you need a clip that loops.
• seed_mode - RANDOM / LOCKED / REUSE LAST, so you can iterate on one composition or chase a seed that worked.

There's also an optional last_frame (IMAGE) for end-frame targeting - feed H3 a destination frame for A→B motion control, the capability the community immediately compared to Kling.

What comes out

A pile of typed outputs (width, height, duration, frame_count, seed, fps, mode flags) plus two things that matter structurally: render_config (a VELVET_VICE_H3_RENDER_CONFIG bundle that drives the rest of the graph) and a human-readable summary/status you can surface on screen. The point of all those passthroughs is that every downstream node reads its settings from one authoritative place instead of a dozen duplicated widgets.

Install and reality check

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

or ComfyUI Manager → velvet-vice-minimax-h3, then restart. The Director itself is free of model downloads - but the render it drives is not. H3 is the 33B open-weights MiniMax model (~42.5GB full precision, or GGUF quants that shrink it), and its Community License excludes the US, EU, UK and South Korea from running the local weights. Check that before you invest an afternoon; the license is the single most common reason someone discovers H3 after they can't run it.

Where beginners burn time: they set turbo_enabled on with no Turbo LoRA installed and wonder why the render is slow - the node only switches step counts when the LoRA is detected. And they ignore the frame-grid snapping, then complain the clip is 4.7s instead of 5. Both are "read the tooltip" problems, and this node's tooltips are unusually good.