This repository provides PyTorch implementations for objective diffusion models, enabling users to generate images from noise using various sampling techniques. It is designed for researchers and practitioners interested in state-of-the-art generative models, offering flexibility in sampling methods and model conditioning.

How It Works

The project implements denoising diffusion probabilistic models that are trained to reverse a noising process. It utilizes the 'v' objective from Progressive Distillation for faster sampling and supports CLIP-guided diffusion, allowing generation conditioned on text embeddings. The code includes multiple sampling methods like DDPM, DDIM, PRK, and PLMS, offering trade-offs between speed and sample quality.

Quick Start & Requirements

• Install via pip: pip install v-diffusion-pytorch or clone and run pip install -e .
• Requires PyTorch. GPU with CUDA is recommended for performance.
• Official Colab notebook available for a simplified experience.

Highlighted Details

• Supports classifier-free guidance for improved text-conditional generation.
• Offers various sampling methods including DDPM, DDIM, PRK, PLMS, PIE, PLMS2, and IPLMS.
• Includes pre-trained models for different datasets and resolutions (e.g., CC12M, YFCC, Danbooru, ImageNet, WikiArt).
• Provides scripts for both unconditional (cfg_sample.py) and CLIP-guided (clip_sample.py) sampling.

Maintenance & Community

Developed by Katherine Crowson and Chainbreakers AI. Compute resources for training were provided by stability.ai.

Licensing & Compatibility

The repository does not explicitly state a license in the README. Users should verify licensing for commercial or closed-source use.

Limitations & Caveats

The README does not specify a license, which may impact commercial adoption. Some sampling methods might require significantly more steps for optimal results.