MiniMax H3 AV Latent Separate / 联合潜空间拆分 (EXP/T8)

Sometimes you don't need pixels or sound at all - you need to move a latent around. Say you're routing video and audio streams through different branches, or replacing one stream and keeping the other, or just inspecting what a joint H3 latent actually contains. Decoding to do that is a waste of VRAM and time. MiniMaxH3AVLatentSeparateT8 splits the joint latent into video_latent and audio_latent without invoking either VAE, and it's marked EXP for a reason: it's a sharp tool for people who think in latent space.

The mechanism is simple and cheap. H3's joint latent is a nested tensor structure holding both streams plus their per-stream noise masks and metadata. This node validates that structure (it'll complain if what you fed it isn't a proper H3 AV latent), separates the streams, and passes the metadata and masks through intact. No VAE encode, no VAE decode - just a split on the latent's internal layout. The README's framing for the whole latent family is that these nodes are cheap precisely because they never round-trip through pixels.

The one input, the three outputs

One input, av_latent. Three outputs:

• video_latent - the video stream alone, ready to route or replace.
• audio_latent - the audio stream alone.
• report - a STRING telling you what the node validated and how it split things. Feed it to a text node.

The killer use case is stream replacement: separate the latent, swap in a different audio latent (or video latent), recombine, sample again - all without ever touching a VAE. That's dramatically cheaper than the decode → re-encode round trip, which is exactly what the node description promises: "cheaper than decoding when only latent routing or replacement is needed."

Where it fits

It lives in T8/MiniMax H3/Source AV/Experimental, a category of nodes that operate directly on source audio-video material in latent space. It's a sibling to the source-media windowing and preparation nodes in the same menu. It is not a substitute for MiniMaxH3AVDecodeT8 - if your downstream needs IMAGE frames or AUDIO tensors, decode. This one exists for the latent-to-latent paths.

Caveats, honestly: it's experimental, so expect the contract to be stricter than stable nodes (if the latent shape is off, it errors rather than guessing), and there's no recombine node in this article's scope - you recombine by feeding the separated pieces back into whatever node builds the joint latent, or by working with the pack's native latent timeline tooling. Start from the 04-long-video workflows if you want to see the latent-native approach in action.

Install

Same pack install as everything else from T8mars:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart ComfyUI. No extra pip dependencies, no model files - it touches only the latent tensor, which is why it's one of the cheapest nodes in the pack to run.