OpenAI open-sourced DALL-E's eyes, but not its brain

What it does

This package implements the discrete variational autoencoder from the original DALL-E paper in PyTorch. It encodes images into a compact grid of discrete tokens and decodes them back into pixels. The text-conditional transformer that actually generates images from captions is not included, so this is strictly the compression half of the pipeline.

The interesting bit

The repo releases the official “perceptual backbone” — the component that turns pixels into a vocabulary a transformer can read. What makes it notable is the asymmetry: OpenAI shipped the image tokenizer as a standalone tool, yet kept the generative text-to-image model closed, leaving you with a very well-engineered compressor but no way to summon a cat in a hat from a prompt.

Key highlights

• Official PyTorch port of the DALL-E discrete VAE
• Compresses images to and from a discrete latent token grid
• Explicitly excludes the text-to-image transformer
• Over 10,800 GitHub stars for a partial pipeline release
• README is minimal; usage is demonstrated in a linked notebook

Caveats

• You cannot generate images from text; you can only encode or decode existing images
• The README offers almost no detail beyond installation and a notebook link
• It is a partial release: the transformer that drives generation is missing

Verdict

Researchers building custom generative models or studying discrete VAEs get official weights and a clean reference implementation. Anyone hoping to type a prompt and receive a picture will leave empty-handed.