MiniMax H3 • Replace Audio Stream (deprecated)

The display name says it plainly: "Replace Audio Stream (deprecated)." The author kept this node around for one reason - so the hundreds of saved graphs that reference it keep loading. The actual node description spells out the rule: use "Replace Stream" for new graphs, with stream=audio. This is the legacy alias that locked the stream choice to audio.

What it still does

If you load an old workflow, this node behaves exactly like the current Replace Stream with stream=audio:

• av_latent - the H3 AV latent whose audio stream you're swapping.
• replacement - the new audio latent (encode it with Encode Audio Stream first).
• fit_mode (strict / crop_pad) and alignment (start / end / center) - how the replacement is fitted. strict errors on length mismatch; crop_pad + alignment handles it.
• denoise (default 0) - 0 preserves the replacement exactly; 1 fully denoises it. For swapping in your own soundtrack, leave it at 0 - you want the source audio untouched, not "improved."

One output: the modified av_latent.

So the capability is real: replace the audio stream in an H3 AV latent, which is the load-bearing step for "keep the generated video, use my audio." The deprecation is about naming and consolidation, not function.

Why it was deprecated

The pack merged the per-stream replace nodes (Replace Video, Replace Audio) into a single Replace Stream node with a stream selector. One node, one schema, two streams. Keeping both old classes registered means old workflows deserialize without the "missing node" red box - but every new graph should use Replace Stream. There's no functional advantage to the old node, and the deprecated name is a hint that future releases may stop shipping it.

Install

Ships in ComfyUI-MiniMax-H3-LongMedia. Manager search "MiniMax-H3-LongMedia", or:

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-MiniMax-H3-LongMedia

restart. H3 checkpoint + VAEs (MiniMaxAI/MiniMax-H3, ~42.5 GB, community license excludes US/EU/UK/Korea) are needed for the workflows around it.

Troubleshooting

If the replacement seems to have no effect, check the denoise default of 0 is doing what you think - it preserves the replacement, so if the output audio still sounds like the model's generated track, you may have wired the wrong latent in (or the sampler re-generated audio after this node ran, which happens if the plan's audio mode is set to generate). And if you're starting a new workflow, just use Replace Stream instead - you'll get the same behavior with a stream=audio selector and a display name that doesn't say "deprecated" on every run.