Replicate andreasjansson/deepfloyd-if

DeepFloyd IF was, for a while, the model that could actually write words in an image and follow a prompt closely - back before Ideogram and the modern crowd. It's architecturally unusual and historically important, and this node lets you generate from it through Replicate's cloud with a one-line prompt. Practically, it's a fixed-quality "type a prompt, get an image" generator: the node exposes almost no knobs, so you're using it for its distinctive output, not for fine control.

Standard Replicate rules: the model runs on Replicate's GPU, not yours. You need a Replicate account and a REPLICATE_API_TOKEN, you pay per run, and nothing installs locally beyond the wrapper.

How it works

Two things make DeepFloyd IF different. First, it's a pixel-space model - it diffuses directly on pixels with no VAE, where nearly every model since SD 1.5 works in a compressed latent and decodes with an autoencoder. Second, it uses a large T5 text encoder and generates in a cascade (a small base image, then super-resolution stages up to 1024px). The T5 encoder is why it reads full sentences and renders text well; the pixel-space cascade is why it's heavy and slow. It was ahead of its time and its ideas resurfaced years later in the 2026 wave of no-VAE models.

The inputs and outputs that matter

This node keeps it minimal - really just:

• prompt - a full natural-language description. T5 handles sentences well, so write like you mean it, and put any text you want rendered in quotes.
• seed - fixes the result so you can reproduce or iterate.
• force_rerun - bypass Replicate's cache when identical inputs would return the same image.

Output is a single IMAGE you wire into a preview or save node. There's no step count, guidance, or size control exposed here - the cascade runs its fixed pipeline.

How to install it

One install brings every Replicate node.

• ComfyUI Manager: search ComfyUI-Replicate, install, restart.
• Manual:

  cd ComfyUI/custom_nodes
  git clone https://github.com/replicate/comfyui-replicate
  cd comfyui-replicate
  pip install -r requirements.txt
  

  then restart.

No weights download - DeepFloyd runs on Replicate. Only the replicate pip package installs. Set your token first: export REPLICATE_API_TOKEN="r8_..." (replicate.com/account/api-tokens).

Common issues & troubleshooting

There are barely any settings. Correct - this node exposes just the prompt and seed. If you want to tune steps, guidance, or resolution, DeepFloyd through this wrapper isn't the place; use a Flux or SDXL node in this same pack for that kind of control. Here, you shape the output almost entirely through the prompt.

It's slow compared to other nodes. Pixel-space diffusion has compute that scales with pixel count, and DeepFloyd runs a multi-stage cascade on top - that's inherent to the architecture, not a Replicate hiccup. It was never a fast model.

The output feels dated next to Flux or Ideogram. It is older, and both of those beat it now on general quality and (for Ideogram) text. Reach for DeepFloyd when you specifically want its look or want to play with a genuinely different, historically important architecture - otherwise the newer nodes in this pack will serve you better.

Same image on a re-run. Replicate cached it for identical inputs - change the seed or set force_rerun.