zlacker

Generating at 64x64px then upscaling it probably gives the model a substantial performance boost (training speed/convergence) than working at 256x256 or 1024x1024 like DALL-E 2. Perhaps that approach to AI-generated art is the future.