DALL-E 1 used a GPT approach to generate both text and images. Images are divided into patches, about 1024 patches for one image. Each patch is like a text toke

DALL-E 1 used a GPT approach to generate both text and images. Images are divided into patches, about 1024 patches for one image. Each patch is like a text token. > We describe a simple approach for this task based on a transformer that autoregressively models the text and image tokens as a single stream of data. https://arxiv.org/abs/2102.12092 https://arxiv.org/abs/2102.12092 The moral - you can just stream together text and images into a GPT style model.