沐阳 H3 · Latent 透传（内部）

It takes a LATENT and returns that same LATENT. Not a copy, not a clone - the same object, straight through. Two lines of Python, and every line of it is a comment about graph architecture.

Why you'd ship an identity node

The pack's source explains it. The Director's expansion code treats a segment's latent as a node with output 0. That's the contract: to the dynamic graph, a segment latent comes from a node, you take out(0), and everything downstream is built around that assumption.

Continuous-Sigma sampling breaks it. That path returns two things from one sampler - the noisy split-state and the clean x0 preview - so the thing the Director wants is out(0) of a node that returns a tuple, and the expansion code would have to branch in every place that touches a segment latent.

H3LatentIdentity restores the contract instead of branching. Wire the sampler's latent through it, take the identity node's samples, and the rest of the pipeline never has to know that the value arrived from a two-output sampler. The comment in the source is unusually candid about this: preserves that stable contract without branching the whole pipeline.

That's a legitimate and common design move in large packs, and it's the same family as the null-output placeholders and fallback switches in the wider ecosystem - the node plumbing layer exists precisely because this kind of invisible node is where real packs hide their architecture.

Inputs and output

• samples (LATENT, required).
• Output samples (LATENT) - the input.

No widgets, no settings, nothing to tune. It's category 沐阳 H3/导演台/内部.

What it costs you

Nothing measurable. There's no tensor operation, no device transfer, no copy - the dict goes through by reference. It is free enough that the pack leaves it in the graph in configurations that don't need it, rather than conditionally building it.

The one thing to be careful about when you're hand-editing a graph: because it does nothing, it's easy to assume it's junk from an older version and delete it. Sometimes it is genuinely disposable. Sometimes the Director's expansion code downstream is reaching for out(0) of that exact node and your graph quietly stops matching what the pack expects. If you're editing a Director-generated graph, leave the boring nodes alone until the thing runs.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang

Restart and hard-refresh; it's under 沐阳 H3/导演台/内部. Or search "ComfyUI MiniMax H3 Myang" in ComfyUI Manager. No Python dependencies - the pack declares an empty dependency list, and this node doesn't even import beyond what the module already has.

You'll still need your own H3 diffusion model, Qwen text encoder, video VAE and audio VAE to do anything with it. The pack ships the workflow, not the weights, and MiniMax H3's community licence is worth reading before you download any of it - it's territory-restricted, and that's a licensing question rather than a technical one.