vlo Set Audio Latent Binary Masks

Modern open video models don't just draw a picture - LTX and MiniMax H3 generate audio with it, jointly, from a shared AV latent. That's great until a generation nails the video and mangles one section of the sound, or you want to regenerate just the dialogue in a region while keeping everything else. Video inpainting has a clean mental model (mask the pixels, resample). vloSetAudioLatentBinaryMasks is the equivalent for the audio stream: it stamps a temporal binary noise mask onto an audio latent, so the sampler re-rolls the sound inside the masked time range and leaves the rest alone.

The stock SetLatentNoiseMask handles video masks. Audio has no such standard primitive, and the layouts are model-specific, which is exactly why this node exists: it resolves the audio latent's layout and rate from the latent and VAE metadata itself, and knows the two current families - LTX's [B,C,T,F] layout and MiniMax H3's [B,C,S,T], with 25 Hz audio latent rate for LTX and 40 Hz for MiniMax.

What you feed it

• audio_latent - a standalone audio latent, or a nested AV latent (a joint H3 or LTX one). If it's AV, the video mask is preserved; only the audio mask changes.
• masks - binary mask frames. Each frame is reduced to active/inactive via threshold (default 0.5, "any pixel meets it"): the spatial content of the mask is irrelevant, only when it's active. Think of it as drawing an audio automation envelope, not a region.
• mask_fps - the FPS your mask frames were made at. Above zero, mask timestamps map onto the VAE's audio latent rate; zero (default) stretches the whole mask batch across the full audio latent length.
• resize_mode - nearest (default) keeps hard frame ranges; linear smooths the timeline before the final binary threshold.
• existing_mask_mode - overwrite (default), add (takes the max with any existing mask), or subtract (clears regions out of an existing mask). Subtract is how you carve a retake window out of a mask you built earlier.

The last three are convenience/advanced: layout_override (auto prefers metadata, then sniffs the VAE), audio_latent_rate, and an optional audio_vae that lets the node resolve rate and layout automatically when the latent carries no metadata. The output is the same audio_latent type with its noise mask set, ready to feed your sampler.

Where people get burned

Audio latent rates are not a place for vibes. If your masks were authored at a different FPS than you declare in mask_fps, the masked region lands in the wrong part of the timeline - that's the input to double-check first. And on a nested AV latent, don't expect the video mask to ride along unless the node for video masks ran first; this node explicitly changes only the audio stream. It comes from the ComfyUI-vlo pack, which also ships an LTX-only sibling (LTX Set Audio Latent Binary Masks) with a simpler fixed set of inputs - this node is the generalist that adds mask FPS, layout resolution and the MiniMax layout.

Install

Whole-pack install, nothing extra to pip:

cd ComfyUI/custom_nodes
git clone https://github.com/PxTicks/ComfyUI-vlo.git

then restart ComfyUI, or search "ComfyUI-vlo" in ComfyUI Manager (the repo is young enough that the clone path is the reliable one). And the elephant: MiniMax H3's weights are big and carry a licence that excludes the US, EU, UK and Korea - LTX is the consumer-VRAM-friendly route if that's what you have, though its own licence is engine-style (free under $10M/yr revenue), not permissive open source.