bghira/minimax-music-suno-reggae-rank128-v2

About this experiment (failed, informative)

This rank-128 attention-only LoRA was an experiment to answer one question: can MiniMax Music 3 be trained without an RVQ encoder, using plain latent MSE flow loss on real audio?

It cannot — and this run demonstrates why. MiniMax Music's DiT is conditioned on embeddings produced by its autoregressive LM. Without an audio→RVQ-codes encoder, there is no way to produce the conditioning that actually corresponds to a real training clip; the cached LM rollouts used here are sampled from the text/lyrics alone and describe a different realization of the song than the ground-truth latents they were paired with. Trained on those misaligned (conditioning, latent) pairs, the model's surface texture drifts toward the dataset while musical structure degrades — audible in the validation audio as increasingly incoherent output over training.

Training was stopped once the negative result was established rather than carried to the full step budget.

bghira/minimax-music-suno-reggae-rank128-v2

This is a PEFT LoRA derived from MiniMaxAI/MiniMax-Music3.

No validation prompt was used during training.

None

Validation settings

• —CFG: 1.7
• —CFG Rescale: 0.0
• —Steps: 30
• —Sampler: FlowMatchEulerDiscreteScheduler
• —Seed: 42
• —Resolution: 256

Note: The validation settings are not necessarily the same as the training settings.

<Gallery />

The text encoder was not trained. You may reuse the base model text encoder for inference.

Training settings

• —Training epochs: 22
• —Training steps: 2750
• —Learning rate: 1.0
• —Learning rate schedule: constant
• —Warmup steps: 0
• —Max grad norm: 0.5
• —Effective batch size: 8
• —Micro-batch size: 1
• —Gradient accumulation steps: 1
• —Number of GPUs: 8
• —Gradient checkpointing: False
• —Prediction type: flow_matching
• —Optimizer: prodigy
• —Trainable parameter precision: Pure BF16
• —Base model precision: no_change
• —Caption dropout probability: 0.1%

• —LoRA Rank: 128
• —LoRA Alpha: 128.0
• —LoRA Dropout: 0.1
• —LoRA initialisation style: default
• —LoRA mode: Standard

Datasets

suno-reggae

• —Repeats: 0
• —Total number of audio files: 123
• —Sample rate: 48 kHz
• —Channels: 2
• —Duration buckets: 30s (1), 40s (1), 50s (2), 60s (2), 70s (2), 80s (3), 90s (2), 100s (2), 110s (4), 120s (11), 130s (13), 140s (15), 150s (23), 160s (21), 170s (21)
• —Used for regularisation data: No

Inference

import torch
from diffusers import DiffusionPipeline

model_id = 'MiniMaxAI/MiniMax-Music3'
adapter_id = 'bghira/minimax-music-suno-reggae-rank128-v2'
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) # loading directly in bf16
pipeline.load_lora_weights(adapter_id)

prompt = "An astronaut is riding a horse through the jungles of Thailand."
negative_prompt = 'blurry, cropped, ugly'

## Optional: quantise the model to save on vram.
## Note: The model was not quantised during training, so it is not necessary to quantise it during inference time.
#from optimum.quanto import quantize, freeze, qint8
#quantize(pipeline.transformer, weights=qint8)
#freeze(pipeline.transformer)
    
pipeline.to('cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu') # the pipeline is already in its target precision level
model_output = pipeline(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=30,
    generator=torch.Generator(device='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu').manual_seed(42),
    width=256,
    height=256,
    guidance_scale=1.7,
).images[0]

model_output.save("output.png", format="PNG")