monster diffusion

Generates 48x48 images of monsters.

FID@20	FID@100
52.5	41.8

FID between train and early_stopping is 7.36. FID@k means that the samples require k model evaluations.

Inspired by:

• DiVAE : Photorealistic Images Synthesis with Denoising Diffusion Decoder
• DiffuseVAE: Efficient, Controllable and High-Fidelity Generation from Low-Dimensional Latents
• Elucidating the Design Space of Diffusion-Based Generative Models
• Katherine Crowson's k-diffusion
• Score-Based Generative Modeling through Stochastic Differential Equations
• Velocity diffusion sampling
• Progressive Distillation for Fast Sampling of Diffusion Models
• Diffusion Models Beat GANS on Image Synthesis

Install

poetry add git+https://github.com/samedii/monster-diffusion.git

Usage

import monster_diffusion

image = monster_diffusion.sample()

Development

Installation

Setup environment:

poetry install

If you need cuda 11:

poetry run pip uninstall torch torchvision -y && poetry run pip install torch==1.11.0 torchvision==0.12.0 --extra-index-url https://download.pytorch.org/whl/cu113

Debugging

Add a debug config for a python module in vscode, e.g.:

{
  "version": "0.2.0",
  "configurations": [
    {
      "name": "Python: Module",
      "type": "python",
      "request": "launch",
      "module": "operations.debug.train",
      "justMyCode": true
    }
  ]
}

Also set a shortcut like Shift+Enter for "Evaluate in debug console" to interactively run code you are trying to fix while debugging.

Training

guild run train
guild run retrain model=<model-hash>
guild run evaluate model=<model-hash>
guild tensorboard <model-hash>