mimetic-a0 Decode

Every latent-to-audio node in ComfyUI looks the same until you listen to what comes out of it. This one turns the sampler's 64-channel latent back into 48 kHz stereo, and then makes one decision you should actually care about: how loud the result is.

The pack is blunt about why it exists. The core VAE Decode Audio node rescales volume; this one doesn't, and instead offers you a straight peak normalise. That difference is the whole point of the node, and it's the difference between a take you can drop into an edit and a take that needs another pass in an audio editor.

What it does

It decodes each take separately - one VAE call per item in the batch, then concatenated. On a four-take batch that's deliberately slower than one big decode and much kinder to your VRAM, which is the right trade for a batch where you're going to throw three takes away.

Then it trims. The sampler generated six seconds of extra window (up to a point) and stashed the real length inside the latent; this node reads that and cuts the audio down to exactly the seconds you asked for. So the trimming only happens on a latent that came out of this pack's Sampler - hand it a latent from elsewhere and there's nothing to trim against.

Finally the checkbox. normalize on, and each take is divided by its own peak and scaled so it lands at −1 dBFS, which is what the command line's WAV files do. Off, and you get the raw output - which, per the node's own tooltip, can go past full scale.

That's the whole surface. samples and vae in, audio out.

The inputs, briefly

samples is the LATENT. vae is the VAE output on the mimetic-a0 Loader - and it has to be that one. The latent space here is the 48 kHz ACE-Step 1.5 autoencoder layout, downsampling 1920, so a VAE borrowed from a Stable Audio checkpoint is decoding a different language. You'll get noise, hiss, or a very quiet smear rather than an error message, which is the annoying kind of failure.

normalize defaults to on, and for most people most of the time that's correct. Two cases where it isn't:

• You're layering parts. Normalise four takes of different elements - kick, bass, chords, hat - and they all come out equally loud at −1 dBFS. Sum those and you clip instantly; you'll be pulling every track down before you can hear whether the arrangement works. With normalize off you keep the model's own relative levels between takes, which is closer to a mix already in progress.
• You're judging takes by ear against each other. Peak normalising makes a thin, quiet take just as loud as a good one. It won't make it better.

The other side of the argument: leave it off and a hot take clips on save, and a quiet one is a whisper in your timeline. Normalising is a safety rail. Turn it off once you know why.

One more consequence of per-take normalising, since it's easy to trip over: whatever's quietest in that take - the noise floor, the tail of a reverb - comes up by the same factor as the drums. The audio here is autoencoder-decoded General-MIDI-flavoured samples, so it isn't pristine to begin with, and boosting a quiet take won't do it favours.

Wiring it up

audio goes into a Save Audio node or an audio preview. The waveform arrives on the CPU as float32 at 48 kHz, stereo - so keep an eye on the bit depth if you're going to do anything else with it. A 16-bit WAV is a fine delivery format and a poor intermediate.

The full chain, top to bottom, is five nodes plus core: mimetic-a0 Loader → mimetic-a0 Prompt → CLIP Text Encode → mimetic-a0 Sampler → this node → Save Audio. If you'd rather see a working graph than a description of one, workflows/gui/text_to_music.json is exactly that, and four_takes.json shows the batched version.

Install

It comes with the pack - Manager (search the pack title), or:

cd ComfyUI/custom_nodes
git clone https://github.com/hiroki-abe-58/ComfyUI-MimeticA0
python custom_nodes/ComfyUI-MimeticA0/tools/download_model.py --models-dir models

That second command puts the weights in models/mimetic_a0/mimetic-a0/ and t5_base.safetensors in models/text_encoders/, both pinned and SHA-256 verified. Nothing is fetched at run time, and there's nothing in requirements.txt to install - the pack runs entirely on ComfyUI's own stable-audio classes and torch. Restart ComfyUI once after the clone and once after the download.

Last thing, and it's the one that bites: with normalize off, check the take's peak before you save a 16-bit WAV, because the raw output can exceed full scale and a 16-bit file clips hard rather than politely. Float WAV or FLAC in the meantime, then normalise deliberately instead of by accident.