DeepFloyd IF
Stability AI's cascaded pixel-diffusion model that finally made text render legibly inside generated images.
About
DeepFloyd IF, released by Stability AI's DeepFloyd research lab in May 2023, was a deliberate departure from latent diffusion: a three-stage cascaded pixel diffusion model (64x64 base, then 256x256 and 1024x1024 super-resolution stages) conditioned on the frozen T5-XXL-1.1 text encoder. The T5 encoder is what gave it the era's best prompt understanding, and its signature trick was rendering legible text and logos inside images, the weakness every other model had at the time. The HF model card reports a zero-shot FID of 6.66 on COCO, and the base UNet is among the largest ever used in a first cascade stage.
It launched under the DeepFloyd IF License, a non-commercial research license, with Stability saying a fully open version would follow, it never did. The weights require gated access, and the pipeline needs roughly 16GB+ of VRAM. DeepFloyd IF also does zero-shot image-to-image translation by noising the input down to 64px and denoising under a new prompt. NightCafe got early access and it was the model behind the era's text-in-image memes. Its ideas (T5 conditioning, cascaded pixel stages) were absorbed into later models, but the model itself is a research artifact now.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Zero-shot FID 6.66 on COCO (author-reported on the HF model card)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (verified tier, metrics-v1). Evidence: DeepFloyd/IF-I-M-v1.0 HF model card (cascaded pixel diffusion, T5-XXL encoder, zero-shot FID 6.66); TechCrunch and MarkTechPost launch coverage confirming the non-commercial research license and the eventual-open-source plan. Strengths: landmark text-rendering capability, published FID number, full architecture released. Docked for: (1) non-commercial license and the promised fully-open release never materialized; (2) heavy VRAM needs and no active maintenance, so it is a research artifact rather than a tool. License: DeepFloyd IF License (non-commercial).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · DeepFloyd / Stability AI; released 2023-05
- · Cascaded pixel diffusion (64->256->1024), T5-XXL-1.1 encoder
- · Zero-shot FID 6.66 on COCO (author-reported on HF model card)
- · Promised fully-open release never materialized; gated weights