mimetic-a0 Loader

mimetic-a0 is a small text-to-music model by Mick Grierson - MIT licensed, trained only on CC0 Freesound clips and synthetic audio, writing 48 kHz stereo in chunks up to 16 seconds. Drum loops, basslines, chords, one-shots, whole little mixes. It is not a Suno competitor; it is the model you reach for when you want a four-bar loop to sit under a video, or a sample you can chop.

The Loader is where that starts, and it's also where most people stall out, because this pack needs files in two different models/ subfolders and only one of them is the "model."

What it actually loads

One node, two outputs, because the thing you're loading is two things. model is the diffusion transformer - ComfyUI's own stable-audio DiT class at 1024 wide, 16 layers, 16 heads, running rectified flow. vae is the audio autoencoder, which here is the 48 kHz ACE-Step 1.5 layout (the "Oobleck" autoencoder, downsampling of 1920 samples per frame). That second output is why you don't add a separate VAELoader to your graph: the VAE ships inside the same weight file and this node hands it to you.

There's a wrinkle worth understanding, because it explains why no other loader will open this model. When ComfyUI sees a stable-audio-ish checkpoint it assumes several things: v-prediction, a specific DiT shape, and length embedders that cover 0–512 seconds. mimetic-a0 breaks all three. It's rectified-flow, its DiT is smaller, and its two number embedders only cover 0–32 seconds. So the Loader reads model_config.json from the model folder, verifies the release really is this model (objective, sample rate, sample size, downsampling ratio), detects the UNet config from the weights, cross-checks the width and depth, and pulls the head count out of the config because the weights don't store it. You can't do that with a stock checkpoint loader, so don't try.

The two inputs that matter

model is a dropdown of subfolders in models/mimetic_a0/, and only folders that contain both model_config.json and a weight file show up in it. That means the layout is models/mimetic_a0/mimetic-a0/, not a pile of safetensors sitting loose in models/mimetic_a0/.

weights picks between fp32 (model.safetensors, 2 GB) and fp16 (model.fp16.safetensors, 1 GB). fp32 is what the author's verification against the official code was run on, and the whole model plus T5 peaked around 3.8 GB of VRAM in their testing, so on anything modern just leave it on fp32. fp16 is for when you're memory-tight or impatient, and you should know the trade: the official code keeps its sampler state in fp16 too, ComfyUI keeps fp32, so a seed no longer gives you the same take as the command line - or as the fp32 file.

Wire model into the Sampler's model, and vae into the Decode node's vae.

Install

Manager, the Comfy Registry id mimetic-a0-music, or the usual:

cd ComfyUI/custom_nodes
git clone https://github.com/hiroki-abe-58/ComfyUI-MimeticA0

There is nothing to pip install - requirements.txt in the repo is a comment saying exactly that, and the declared dependencies list is empty. Given how many audio packs break three other models' transformers version on install, that's genuinely refreshing.

Weights come down with the pack's own script, which pins the Hugging Face revisions and checks every file against a SHA-256:

python custom_nodes/ComfyUI-MimeticA0/tools/download_model.py --models-dir models
# add the fp16 weights too:
python custom_nodes/ComfyUI-MimeticA0/tools/download_model.py --models-dir models --weights all

That fills models/mimetic_a0/mimetic-a0/ and drops t5_base.safetensors into models/text_encoders/. If you already have a t5_base.safetensors there, it's kept rather than overwritten. Restart ComfyUI, then open one of the two example workflows in workflows/gui/.

Common issues

The model dropdown is empty. The folder structure is wrong, or the script never ran. It needs model_config.json next to the weights, one folder deep.

no mimetic-a0 release named ... in models/mimetic_a0/. Same cause, said out loud. The error even points you back at the download command.

does not hold the mimetic-a0 release (model_config.json differs). You pointed it at some other stable-audio-shaped folder. This loader is deliberately picky rather than silently producing noise.

You picked fp16 and it can't find a file. The default download pulls fp32 only. Run the script again with --weights all, or with --weights fp16.

The prompt sounds like it's being ignored. Different node, classic mix-up: this Loader does not load the text encoder. You still need a core Load CLIP set to t5_base.safetensors with type stable_audio, feeding a CLIP Text Encode into the Sampler. Two files, two folders - one of them is just not in the Loader.

Models live outside ComfyUI/models? Add a mimetic_a0 key to extra_model_paths.yaml and the dropdown will pick those up as well.