VELVET VICE MiniMax H3 — Audio Decode / Mute Gate

MiniMax H3 doesn't add audio in a second pass - the soundtrack is generated with the picture and lives in a packed AV latent. That's great when you want sound, and a waste of a GPU when you don't. VelvetViceMiniMaxH3AudioDecodeGate is the node that makes "no audio" actually cheap: when you're rendering video-only it never runs the audio VAE decode at all, instead of running it and throwing the result away.

This is an internal node from the Velvet Vice MiniMax H3 pack (velvet-vice-minimax-h3). In the reference workflow you don't wire it by hand - the WITH SOUND / MUTED · VIDEO ONLY decision from the H3 Director flows into it. But if you're building your own H3 graph, it's the piece that keeps a video-only render honest about memory.

How it works

The node takes the packed H3 latent (samples) and the audio VAE, and its enabled switch decides what happens. Enabled (the default) means "decode": it hands the latent and VAE straight to ComfyUI's own VAEDecodeAudio node and returns real audio. Disabled means the whole branch goes quiet - it returns no audio and never even touches the audio VAE, which is the point. Video-VAE work can't fully overlap on most cards, and decoding a second latent stream just to silence it is a couple of wasted seconds plus a big transient spike.

Inputs and outputs

• samples (LATENT) - the packed H3 AV latent straight out of the sampler.
• vae (VAE) - the native H3 audio VAE, not the video one. Wiring the video VAE here produces garbage that no amount of retrying fixes.
• enabled (BOOLEAN, default true) - the mute gate itself. False = video-only, no decode, no audio.

Output is a single audio (AUDIO) value that you feed to whatever encodes the final file - in this pack that's the H3 Output Studio's audio input.

One honest gotcha: if enabled is on but no audio VAE is connected, the node raises - "WITH SOUND requires a loaded native H3 audio VAE." It's deliberately strict so a silent video never sneaks out labeled as having audio. If you're getting that error and don't care about sound, flip enabled off instead of hunting for the VAE.

Installing and using it

The pack is a ComfyUI custom-node install, no Python extras (its pyproject.toml lists zero dependencies - everything here leans on ComfyUI's own native H3 support):

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

or search for velvet-vice-minimax-h3 in ComfyUI Manager and restart. You still need the actual H3 weights (diffusion model, Qwen3-VL text encoder, video and audio VAEs) from HuggingFace, plus - outside a few territories - the license to run them at all.

The practical rule: leave it wired to the Director's audio toggle and only touch enabled directly when you want a video-only render from a graph that otherwise assumes sound. Audio-off means the audio VAE stays unloaded for the whole run, which on an H3-sized workload is more VRAM headroom than it sounds.