https://github.com/josefalbers/aggressor

Ultra-minimal autoregressive diffusion model for image generation
https://github.com/josefalbers/aggressor

artificial-intelligence audio-generation autoregressive autoregressive-diffusion autoregressive-models deep-learning diffusion diffusion-models image-generation mlx tts

Last synced: 4 months ago
JSON representation

Ultra-minimal autoregressive diffusion model for image generation

• Host: GitHub
• URL: https://github.com/josefalbers/aggressor
• Owner: JosefAlbers
• License: apache-2.0
• Created: 2024-09-12T23:07:52.000Z (almost 2 years ago)
• Default Branch: main
• Last Pushed: 2024-10-06T04:03:56.000Z (almost 2 years ago)
• Last Synced: 2025-12-22T00:38:05.707Z (7 months ago)
• Topics: artificial-intelligence, audio-generation, autoregressive, autoregressive-diffusion, autoregressive-models, deep-learning, diffusion, diffusion-models, image-generation, mlx, tts
• Language: Python
• Homepage:
• Size: 2.36 MB
• Stars: 21
• Watchers: 2
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Aggressor: Ultra-minimal autoregressive diffusion model for image and speech generation

**CIFAR**

**MNIST**

**AUDIO**

![cifar](https://raw.githubusercontent.com/JosefAlbers/Aggressor/main/assets/aggressor_cifar.png)

![mnist](https://raw.githubusercontent.com/JosefAlbers/Aggressor/main/assets/aggressor_mnist.png)

[audio](https://github.com/user-attachments/assets/89574625-2ec0-4aeb-884f-bf03e5a4aab0)

A simplest possible implementation of [Autoregressive Image Generation without Vector Quantization](https://arxiv.org/abs/2406.11838).

## Key Features

- **Simple Architecture**: A tiny transformer for autoregression and an MLP for diffusion.

- **Minimal Dependencies**: Built from scratch using only basic MLX operations.

- **Single-File Implementation**: Entire model in one Python file `aggressor.py`.

## Components

- `Aggressor`: Main model class combining transformer and diffusion.

- `Transformer`: Multi-layer transformer with attention and MLP blocks.

- `Denoiser`: MLP-based diffusion process with time embedding.

- `Scheduler`: Handles forward and backward processes for diffusion.

## Usage

```zsh

python aggressor.py

```

*(Training on 60000 images x 20 epochs takes approximately 7~8 minutes on 8GB M2 MacBook.)*

## Acknowledgements

Thanks to [lucidrains](https://github.com/lucidrains/autoregressive-diffusion-pytorch)' fantastic code that inspired this project. The official implementation is available [here](https://github.com/LTH14/mar).