Today, Stability AI announced the availability of SDXL 0.9, the most advanced development in the Stable Diffusion suite of text-to-image models. Following the successful release of the Stable Diffusion XL beta in April, SDXL 0.9 significantly improves image and composition detail over its predecessor.

The model is accessible today via ClipDrop, with an API coming soon. Research weights are available now, and the open release will roll out in mid-July, as we move into version 1.0.

Although it is capable of running on modern consumer-grade GPUs, SDXL 0.9 provides a leap forward for creative use cases of generating AI images. Capable of generating hyper-realistic compositions for film, television, music and instructional videos, while also providing advancements for design and industrial use, putting SDXL at the forefront of real-world applications of AI imagery.

example

An example of some hints tested on SDXL beta (left) and 0.9 demonstrates the progress this model has made in just two months.

Prompt: ✨aesthetic✨ aliens walk among us in Las Vegas, scratchy found film photograph

(Left - SDXL Beta, Right - SDXL 0.9)

Prompt: A wolf in Yosemite National Park, chilly nature documentary film photography Negative prompt: 3d render, smooth, plastic, blurry, grainy, low-resolution, anime, deep-fried, oversaturated

(Left - SDXL Beta, Right - SDXL 0.9)

Prompt: *~aesthetic~*~ manicured hand holding up a take-out coffee, pastel chilly dawn beach instagram film photography Negative prompt: 3d render, smooth, plastic, blurry, grainy, low-resolution, anime

(Left - SDXL Beta, Right - SDXL 0.9) The SDXL series also offers a range of features beyond basic text prompts. These features include image-to-image hinting (entering an image to get changes to that image), inpainting (reconstructing missing parts of an image), and extensions (building seamless extensions of existing images).

How does the kernel work?

SDXL 0.9 The significant increase in the number of all weights and biases in the composition composition (the sum of all weights and biases in the neural network the model is trained on) is a key driver of this compositional advancement.

SDXL 0.9 has one of the largest number of parameters of any open source image model, with a 3.5B parametric base model and a 6.6B parametric model ensemble pipeline (the final output is created by running both models and summing the results). The second-stage model of the pipeline is used to add finer detail to the output produced by the first stage.

In comparison, the Beta version runs on 3.1B parameters and uses only one model.

SDXL 0.9 runs on two CLIP models , including the largest trained OpenCLIP model to date (OpenCLIP ViT-G/14), adding 0.9 processing power and the ability to create at higher depth and 1024x1024 higher resolution Realistic image capabilities.

The SDXL team will soon publish a research blog with more details on the model's specifications and testing.

Prompt: beautiful scenery nature glass bottle landscape, purple galaxy bottle (SDXL 0.9 - 1024x1024) System Requirements

Despite its powerful output and advanced model architecture, SDXL 0.9 requires only Windows 10 or 11 or a Linux operating system with 16GB of RAM, a modern consumer grade Nvidia GeForce RTX 20 graphics card (same or higher standard) with at least 8GB of VRAM The GPU is ready to run. Linux users can also use a compatible AMD card with 16GB of VRAM.

Beta Launch Statistics

Since the SDXL beta release on April 13th , we've had a tremendous response from nearly 7,000 users in the digital community. These users have generated over 700,000 images, an average of over 20,000 per day. More than 54,000 images have been entered into the digital community "showcase", of which 3,521 SDXL images were nominated as winners.

Prompt: magical realism; manipulated fingers holding a piece of white heart-shaped sea glass up against the setting sun realistic film photography (SDXL beta - 480x480)

how to get

SDXL 0.9 is now accessible through Stability AI's Clipdrop platform. Stability AI API and DreamStudio customers will also be able to access the model on Monday, June 26, along with other major image generation tools such as NightCafe.

SDXL 0.9 will be available for research use only, for a limited time to gather feedback and fully optimize the model, followed by a full public release. Running code will be publicly released on Github.

If researchers would like access to these models, please apply using the following links: SDXL-0.9-Base Model and SDXL-0.9-Refiner . Please log into your HuggingFace account with your academic email to request access. Keep in mind that currently, SDXL 0.9 is limited to research use.

next step

SDXL 0.9 will follow, with a general public release of SDXL 1.0 expected in mid-July (release time to be determined).

license

SDXL0.9 is released under a non-commercial, research-only license and is subject to its terms of use.