Skip to content

Deep floyd

An open source cascaded text to image model

Visit official site

What Deep floyd is

The model is described as an open source text-to-image system built from a frozen text encoder and cascaded pixel diffusion modules.

Architecture notes

  • Text embeddings feed a UNet architecture
  • Cross attention and attention pooling are added
  • All stages share the T5 based encoder

Output stages

  • The first stage works at 64x64
  • A middle stage reaches 256x256
  • The final stage reaches 1024x1024

Pros & cons

✓ What we like

  • A zero-shot FID score of 6.66 on COCO is stated
  • Text embeddings come from a T5 transformer

! What to watch out for

  • The card describes an open source model
  • The described pipeline is cascaded

FAQ

What resolutions are named?

64x64, 256x256 and 1024x1024 are described.

What score is stated?

The material gives a zero-shot FID of 6.66 on COCO.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review