Generate new latent codes for RAVE with Denoising Diffusion Probabilistic models.
Author: Moisés Horta Valenzuela /
𝔥𝔢𝔵𝔬𝔯𝔠𝔦𝔰𝔪𝔬𝔰
Year: 2023
RAVE-Latent Diffusion is a denoising diffusion model designed to generate new RAVE
latent codes with a large context window, faster than realtime, while maintaining music structural coherency. RAVE-Latent Diffusion can currently be used for unconditional RAVE latent generation, with the smallest context window of aprox. 1:30 minutes of audio (preprocess.py --latent_length=2048 ...
) and can scale up to a context window of 11:30 minutes (preprocess.py --latent_length=16384 ...
).
RAVE-Latent Diffusion can generate audio faster than real-time on a consumer CPU, which allows for wide accesibility.
Audio examples can be found here
, here
and here
.
- Pretrained RAVE model, exported as .ts file:
Make sure that your pre-trained RAVE model has a latent dimension size of powers of two. The default RAVE latent dimension size is 16. For more information on how to modify this, see the original
RAVE
repo. Make sure that when exporting your RAVE model you set--stereo=True
.
- Audio dataset used to train RAVE model:
The recommended audio dataset length is 3 hours, but training on more data will prevent overfitting. You can also train on audio that has not been 'heard' by the pre-trained RAVE model, but ultimately the fidelity of the generated audio will depend on the reconstruction quality that your pre-trained RAVE model has on your audio dataset.
The process of training a new RAVE-Latent Diffusion model consists of three steps:
- Pre-processing by converting your audio into RAVE latents of a fixed context window.
- Training the RAVE-Latent Diffusion model.
- Generate new latent codes with RAVE-Latent Diffusion and decode them into audio with a pretrained RAVE model.
Make a new Python virtual environment or conda environment. Execute the following commands:
git clone https://github.com/moiseshorta/RAVE-Latent-Diffusion.git
cd RAVE-Latent-Diffusion
pip install -r requirements.txt
Code has been tested on Python 3.9.
python preprocess.py --rave_model "/path/to/your/pretrained/rave/model.ts" --audio_folder "/path/to/your/audio/dataset" --latent_length 4096 --latent_folder "/path/to/save/encoded/rave/latents"
This is where you define your context window with the flag --latent_length
. The default value is 4096
latents context window, which is about 2:55 minutes of audio at 48KHz sample rate.
You can adjust the sample rate to match your pretrained RAVE model with the --sample_rate
flag. Currently it supports 44100 and 48000Hz sample rates.
Once you have preprocessed your audio dataset into RAVE latents, you can start training the latent diffusion model:
python train.py --name name_for_your_run --latent_folder "/path/to/saved/encoded/rave/latents" --save_out_path "/path/to/save/rave-latent-diffusion/checkpoints"
Once you have trained the RAVE-Latent Diffusion model, you can now generate new latent codes with it. For example:
python generate.py --model_path "/path/to/trained/rave-latent-diffusion/model.pt" --rave_model "/path/to/your/pretrained/rave/model.ts" --diffusion_steps 100 --seed 664 --output_path "/path/to/save/generated/audio" --latent_length 4096 --latent_mult 1
This command will generate 4096
RAVE latents, defined by --latent_length=4096
and multiply it by 1, defined by --latent_mult=1
. You can generate longer audio by modifying --latent_mult
to 2 or another int. Be aware that your PC might run out of memory if the generated latent is too long.
Keep in mind that RAVE-Latent Diffusion produces best results when the --latent_length
is equal or higher than the context window used to train the diffusion model. This value is defined by the --latent_length
flag in the Preprocessing step.
You can interpolate between two generated latent codes by setting the --lerp=True
flag. This mode will generate two RAVE latent codes of --latent_length
using two random seeds, slerp between them and decode them using a pretrained RAVE model.
python generate.py --model_path "/path/to/trained/rave-latent-diffusion/model.pt" --rave_model "/path/to/your/pretrained/rave/model.ts" --lerp True --diffusion_steps 100 --seed 664 --output_path "/path/to/save/generated/audio" --latent_length 4096 --latent_mult 1
For more control and reproducibility, you can set the seed of the starting and ending RAVE latent codes with the --seed_a
and --seed_b
flags, respectively.
Coming soon...
Coming soon...
- This code builds on the work done at acids-ircam with
RAVE (Caillon, 2022)
. - The denoising diffusion model is based on the open-source
audio-diffusion-pytorch
library. - Many thanks to Zach Evans from Harmon.ai for helping debug the code.
Coming soon...