import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("keonroohparvar/continuer_pipeline", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]ContinuerPipeline
This repo contains most of the work done for my Master's Thesis. The Continuer Pipeline is a pipeline that utilizes a novel Latent Diffusion model architecture to take a piece of music and extend it by 5 seconds.
The pipeline is implemented at a high level in the continuer_pipeline.py script, and it extends the DiffusionPipeline class from HuggingFace to allow ease of use.
The file structure of this repo is the following:
.
├── legacy # Contains most of the development/attempted methods to get this project working
├── .gitignore # Basic Python .gitignore with custom ignores for local data folders
├── results # Folder with some simple examples of
├── README.md # This file
└── continuer_pipeline.py # The main file that contains the pipeline implementation
My Thesis document describes how this tehcnology works in depth, but at a high level, the Continuer Pipeline simply takes in a waveform and predicts what the next 5-second chunk will sound like. It does this using a novel Latent Diffusion model architecture, and ultimately converts all the waveforms to spectrograms to handle this problem in the image space.
- Downloads last month
- 5