DeepFloyd IF

A frozen text encoder feeds three cascaded pixel diffusion modules in DeepFloyd IF, an open-source text-to-image model built for photorealism and close reading of prompts.

The first module generates a 64x64 pixel image from your text.

Two super-resolution models then raise the result to 256x256 and finally 1024x1024 pixels.

That staged design is the main thing to understand about DeepFloyd IF.

Instead of producing a large image in one pass, the model builds a small draft and sharpens it in steps, so each module has a narrower job.

The text encoder stays frozen, which means the language understanding is fixed while the diffusion stages do the image work.

The code lives on GitHub, so this is a model you run and modify yourself rather than a hosted app with a login.

That fits developers and researchers who want to inspect the pipeline or build on it.

The listing also covers related uses such as image-to-image work, inpainting and upscaling.

If you only want to type a prompt and get a picture, a hosted generator will be less work.