DeepFloyd IF
by DeepFloyd Lab at StabilityAI
Open-source text-to-image diffusion model with a T5 text encoder and three cascaded stages, producing up to 1024x1024 px images.
AI image generators Code libraryOpen source
1. Price: github.com/deep-floyd/IF
github.com/deep-floyd/IFNo quote yet: price is read from the official pages.
2. Commercial use: LICENSE on GitHub
“to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software”
The model and the data it produces may not be used for any commercial or production purposes.
LICENSE on GitHub3. Credit the author: LICENSE on GitHub
“The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.”
No credit on screen. Keep the license notice inside the code you ship.
LICENSE on GitHub4. Agent access: github.com/deep-floyd/IF
An agent can install it by itself with pip install deepfloyd_if.
github.com/deep-floyd/IFLicense checked Oct 24, 2026 against 1 official page. Tap a verdict to see its source. How we checked Report an error
DeepFloyd IF: research text-to-image model, non-commercial
DeepFloyd IF is an AI model from DeepFloyd Lab and Stability AI that draws a picture from a text description, up to 1024 by 1024 pixels. It runs on your own computer with a large NVIDIA graphics card. The code is open, but the trained model may be used only for non-commercial research.
Where you can use it
Pictures from a text description
Write what you want to see, and the model draws it in three steps, from a small picture up to 1024 by 1024 pixels. Research use only.
Restyle an existing picture
Give it a picture and a style such as "origami" or "oil painting", and it redraws the picture in that style.
Enlarge a small picture
Make a small picture four times larger with added detail, even if the model did not draw it.
Repaint one part of a picture
Mark an area of a picture and describe what should be there, and the model fills only that area.
Teach it a new subject
Train it on a few photos of one object or character so it can draw that subject in new scenes, for research.
Compare image models in a study Idea
Use it as a fixed, well-known model next to newer ones when you compare how models follow prompts.
As reported by DeepFloyd Lab at StabilityAI: Zero-shot FID score of 6.66 on the COCO dataset; The largest first-stage model, IF-I-XL, has 4.3B parameters.
DeepFloyd IF popularity
Downloads from pypistats.org, stars from GitHub, checked Nov 19, 2026. How the Gid Score is calculated
How to start
Tell your AI assistant
“For a non-commercial research test only, set up DeepFloyd IF with Diffusers in a Python virtual environment, read my Hugging Face token from the HF_TOKEN environment variable, and run stage 1 and stage 2 from the official README on my prompt.”
Show codeStep 2 or Step 3
Install the Diffusers library and its helpers in a Python environment with PyTorch for your graphics card.
pip install diffusers accelerate transformers safetensors
Log in from Python with your own access token, then run the three stages from the README example.
from huggingface_hub import login
login()
Mistakes to avoid
The model license bans commercial and production use of the model and of data it produces.
The weights are gated; the download fails without an accepted license and a token.
The README says to delete all of them on newer torch.
The code license forbids disabling the inference filters.
License
Modified MIT (code); separate DeepFloyd IF license for weights. The trained model is for non-commercial research only, and even its pictures may not be used for commercial purposes. The code has a modified MIT license that forbids turning off its safety filters.
“to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software” LICENSE on GitHub
“The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.” LICENSE on GitHub
“to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software” LICENSE on GitHub
Non-commercial research copies only, with a copy of the license and the DeepFloyd attribution notice.
“Together with any copies of the Software Products (as well as derivative works thereof or works incorporating the Software Products) that you distribute, you must provide (i) a copy of this License,” github.com/deep-floyd/IF/blob/develop/LICENSE-MODEL
AllowedWith conditionsNot allowedNot stated or not checked yet
Is DeepFloyd IF free for commercial use? The full answer Not legal advice. Check the license page before you ship.
For your agent
License rules tell your agent what it may do with DeepFloyd IF. Paste them into your prompt or AGENTS.md.
- DeepFloyd IF (AI image generators). Open-source text-to-image diffusion model with a T5 text encoder and three cascaded stages, producing up to 1024x1024 px images. Price: free; commercial use with conditions; no credit needed; ship inside your app with conditions.
Connect: PyPI deepfloyd_if.
Rule: Keep the copyright and license notice in all copies of the code you ship.
Rule: Do not remove, disable or circumvent the inference filters (safety checker, watermarker) in the Software.
Rule: Accept the DeepFloyd IF weights license on Hugging Face before downloading the model weights, and check LICENSE-MODEL for their terms.
License: https://github.com/deep-floyd/IF/blob/HEAD/LICENSE (checked 2026-10-24; details: https://agentgid.com/assets/if/)
Agent installs a package or loads it from a CDN.
pip install deepfloyd_ifDeepFloyd is licensed under the DeepFloyd License, Copyright (c) Stability AI Ltd. All Rights Reserved.Alternatives
DeepFloyd IF This page
Pick SDXL when you need an open image model whose weights allow commercial use under its own license.
Pick FLUX.1 [dev] for a newer open-weights model whose license is also non-commercial but lets you use the pictures commercially with some limits.
Pick OpenAI's image model when you want a paid hosted service with no graphics card of your own.
Free open-source, self-hosted AI image generation app with a layer-based canvas and node-based workflows.
Efficient open-weights image generation model built on a single-stream diffusion transformer.
Questions
Can I use the pictures in ads, a shop or a paid app?
No. The model license forbids commercial or production use, and this also covers data the model produces.
How big are the pictures?
Stage 1 draws 64 by 64 pixels, and two more stages enlarge it to 256 and then 1024 pixels.
Do I need an account?
Yes. You need a Hugging Face account, accept the license on the model page and log in locally with a token.
What graphics card do I need?
16 GB of graphics memory for the first two stages and 24 GB with the 1024-pixel stage; Diffusers can run it with about 14 GB.
Must I credit DeepFloyd when I share it?
Yes. When you pass on the model or works that include it, add a copy of the license and the attribution notice it gives.