ddpm celebahq 256

Introduction

The DDPM-CelebAHQ-256 model is designed for unconditional image generation using Denoising Diffusion Probabilistic Models (DDPM). It leverages latent variable models inspired by nonequilibrium thermodynamics to produce high-quality image synthesis. The model achieves notable performance on datasets like CIFAR10 and LSUN.

Architecture

Denoising Diffusion Probabilistic Models use a novel approach combining diffusion probabilistic models and denoising score matching with Langevin dynamics. This model architecture supports a progressive lossy decompression scheme similar to autoregressive decoding. It incorporates discrete noise schedulers such as DDPM, DDIM, and PNDM for inference, each offering different trade-offs between quality and speed.

Training

To train a model using this architecture, you can refer to the official training example available on Hugging Face's GitHub via Google Colab. The training process involves optimizing a weighted variational bound to achieve high-quality image generation.

Guide: Running Locally

1. Installation: Ensure you have the diffusers library installed.

   pip install diffusers
   

2. Load Model: Use the DDPMPipeline (or DDIMPipeline, PNDMPipeline for faster inference) to load the model.

   from diffusers import DDPMPipeline
   model_id = "google/ddpm-celebahq-256"
   ddpm = DDPMPipeline.from_pretrained(model_id)
   

3. Generate Image: Run the pipeline to generate an image from random noise.

   image = ddpm()["sample"]
   image[0].save("ddpm_generated_image.png")
   

For enhanced performance, consider using cloud GPUs from providers like AWS, Google Cloud, or Azure.

License

The DDPM-CelebAHQ-256 model is licensed under the Apache 2.0 License, allowing for both personal and commercial use, modification, and distribution.