openai/image-gpt
OpenAI's image GPT (iGPT) applies GPT-2 transformer architecture to generate and model images.

Velocity · 7d
+0.9
★ / day
Trend
→steady
star history
This repository contains code and pre-trained model checkpoints for image GPT (iGPT), a generative model from the paper ‘Generative Pretraining from Pixels’. The model adapts GPT-2 architecture for image generation by working on pixel sequences encoded with a 9-bit color palette. It supports three model sizes (S, M, L) trained up to 1M iterations, and includes utilities to download pre-trained checkpoints and evaluation datasets including ImageNet and CIFAR-10.