Add converted MiniMax Music 3 audio VAE

README.md

@@ -0,0 +1,20 @@
+---
+library_name: diffusers
+pipeline_tag: text-to-audio
+base_model: MiniMaxAI/MiniMax-Music3
+tags:
+- minimax-music-3
+- audio-vae
+- diffusers
+- simpletuner
+---
+
+# MiniMax Music 3 Audio VAE
+
+This repository contains the MiniMax Music 3 DAV audio autoencoder converted to a Diffusers-style component for SimpleTuner.
+
+The converted component is stored in `audio_vae/` and can be loaded with `MiniMaxMusic3DAV.from_pretrained(repo_id, subfolder="audio_vae")` from SimpleTuner.
+
+This is the continuous waveform autoencoder used for VAECache and waveform decode. It is not the RVQ tokenizer, Qwen3 language model, RVQ depth decoder, or flow transformer from the full MiniMax Music 3 pipeline.
+
+Source weights: `MiniMaxAI/MiniMax-Music3` `dav.pth`.

audio_vae/config.json

@@ -0,0 +1,23 @@
+{
+  "_class_name": "MiniMaxMusic3DAV",
+  "_diffusers_version": "0.38.0",
+  "channel_latent_channels": 64,
+  "decoder_hidden_dim": 1536,
+  "decoder_input_dim": 1024,
+  "encoder_dim": 64,
+  "encoder_latent_dim": 1024,
+  "encoder_rates": [
+    2,
+    4,
+    8,
+    8
+  ],
+  "latent_channels": 128,
+  "sampling_rate": 44100,
+  "upsampling_ratios": [
+    8,
+    8,
+    4,
+    2
+  ]
+}

audio_vae/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea6d2458de8d71e3d8b8210362ab31c547ac3c99bafa53ba004f3751acb5428e
+size 306466152