Diffusers documentation
MiniMaxMusic3Transformer1DModel
MiniMaxMusic3Transformer1DModel
The 2.4B flow-matching Diffusion Transformer of MiniMax Music 3. It denoises 128-channel Flow-VAE audio latents conditioned on the per-frame hidden states of the model’s autoregressive language-model stage, prepending the flow-matching timestep as an extra sequence token (a Stable-Audio-lineage continuous transformer with partial rotary attention and GLU feedforwards).
MiniMaxMusic3Transformer1DModel
class diffusers.MiniMaxMusic3Transformer1DModel
< source >( in_channels: int = 128condition_dim: int = 2048num_layers: int = 36num_attention_heads: int = 32attention_head_dim: int = 64ff_inner_dim: int = 8192rotary_dim: int = 32fourier_embedding_dim: int = 256 )
The flow-matching diffusion transformer of MiniMax Music 3. It denoises Flow-VAE audio latents conditioned on per-frame hidden states produced by the autoregressive language-model stage.
Inputs are 1D latent sequences of shape (batch, in_channels, length). The conditioning signal
(encoder_hidden_states, shape (batch, length, condition_dim)) must already be aligned to the latent timeline —
see MiniMaxMusic3ConditionEncoder. The flow-matching timestep runs from 0 (noise) to 1 (data).
forward
< source >( hidden_states: Tensortimestep: Tensorencoder_hidden_states: Tensorreturn_dict: bool = True )
Parameters
- hidden_states (
torch.Tensorof shape(batch, in_channels, length)) — Noisy Flow-VAE latents. - timestep (
torch.Tensorof shape(batch,)) — Flow-matching time in[0, 1], where 0 is pure noise and 1 is data. - encoder_hidden_states (
torch.Tensorof shape(batch, length, condition_dim)) — Frame-aligned conditioning fromMiniMaxMusic3ConditionEncoder. Pass zeros for the unconditional branch of classifier-free guidance. - return_dict (
bool, defaults toTrue) — Whether to return a Transformer2DModelOutput instead of a plain tuple.