Implementation:Facebookresearch Audiocraft DiffusionSolver

Knowledge Sources	Facebookresearch_Audiocraft Multi-Band Diffusion
Domains	Diffusion, Training
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for training Multi-Band Diffusion models that enhance audio quality by learning to denoise individual frequency bands.

Description

DiffusionSolver extends StandardSolver to implement the training loop for diffusion-based audio enhancement models. It trains a U-Net score model conditioned on codec latent embeddings, using MSE or L1 loss normalized by a reference loss. The solver includes RVM (Relative Volume Mel) evaluation and sample generation with spectral EQ matching.

Usage

Use this solver when training Multi-Band Diffusion models for converting codec tokens to high-fidelity audio.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/solvers/diffusion.py
• Lines: 1-279

Signature

class DiffusionSolver(StandardSolver):
    def __init__(self, cfg: DictConfig): ...
    def run_step(self, idx: int, batch: torch.Tensor, metrics: dict): ...
    def evaluate(self) -> dict: ...
    def regenerate(self, wav: torch.Tensor, step_list=None) -> torch.Tensor: ...

Import

from audiocraft.solvers.diffusion import DiffusionSolver

I/O Contract

Inputs

Name	Type	Required	Description
batch	torch.Tensor	Yes	Audio batch [B, C, T]
cfg	DictConfig	Yes	Hydra configuration

Outputs

Name	Type	Description
metrics	dict	Training metrics including loss, normed_loss
regenerated	torch.Tensor	Denoised audio (from regenerate)

Related Pages

• Principle:Facebookresearch_Audiocraft_Diffusion_UNet_Architecture