Stability AI announces high-performance model 'SDXL 0.9' of image generation AI 'Stable Diffusion', greatly improving image and composition details
Stability AI, an AI research company that developed the image generation AI `` Stable Diffusion '', released a beta version of `` Stable Diffusion XL (SDXL) '' in April 2023, which greatly enhanced the conventional image generation AI. And on June 22nd, local time, a new model ' SDXL 0.9 ' with greatly improved image and composition details was released from the SDXL beta version.
Stability AI has released SDXL 0.9, the most performant
#StableDiffusion to date.
Please refer to this blog for details. https://t.co/FMNtygCV4Opic.twitter.com/gOFUiPcf0M _ — Stability AI Japan Official (@StabilityAI_JP) June 22, 2023
Stability AI Releases SDXL 0.9: Breakthrough in Image Generation AI Technology — Stability AI Japan — Stability AI Japan
https://ja.stability.ai/blog/sdxl-09-stable-diffusion
The beta version of “SDXL”, which Stability AI announced in April 2023, has increased the number of model parameters from the previous 900 million to 2.3 billion, and the quality of the images that can be generated has greatly improved. A review of actually generating images using the beta version of SDXL can be read below.
A beta version of the image generation AI ``Stable Diffusion XL'', which is significantly stronger than the conventional Stable Diffusion, has been released as a test, so I tried using it-GIGAZINE
According to Stability AI, the beta version of SDXL received a great response from the Discord community of nearly 7,000 people, and the total number of images generated by Discord users exceeded 700,000.
On June 22, Stability AI announced the latest version 'SDXL 0.9', which has further improved image and composition details from the beta version of SDXL. In SDXL 0.9, the number of parameters has further increased from the beta version, and it consists of a single model with 3.5 billion parameters and multiple models with 5.8 billion parameters, making it one of the largest open source models. By combining two models to generate an image, Stability AI explains that finer details can be added to the image.
On their official blog, Stability AI publishes the results of generating images from the same prompt in the previous model SDXL beta version and the latest model SDXL 0.9. Below are images in SDXL Beta (left) and SDXL 0.9 (right) from the prompt 'aesthetic aliens walk among us in Las Vegas, scratchy found film photograph' is the result of generating Images with SDXL 0.9 have more detail.
Below is the result of generating the prompt 'A wolf in Yosemite National Park, chilly nature documentary film photography'. Compared to SDXL Beta (left), SDXL 0.9 (right) has a more pronounced wolf face.
At the time of writing the article, SDXL 0.9 can be accessed from
After visiting ClipDrop's official website, enter the prompt in the central input field. This time, enter the prompt 'Horse show jumping in nature in good weather' and click 'Generate'.
The image was generated in about tens of seconds. The output result screen has space to display 4 images, but it seems that not all of them are displayed due to server congestion. Click the image to see more.
Although there is a sense of incongruity in the details when looking closely, a fairly high quality image was generated.
As a comparison, let's generate an image from the same prompt using the conventional version of 'Stable Diffusion v2.1 (version 2.1)' in '
The generated image looks something like this.
I enlarged one of them. It follows the prompts, but the details are a little rough compared to the images generated by SDXL 0.9.
When the image (left) generated with Stable Diffusion v2.1 and the image (right) generated with SDXL 0.9 are arranged side by side, it looks like this.
SDXL 0.9 runs on Windows 10/11 and Linux and can be run on a standard PC with an Nvidia GeForce RTX 20 graphics card with 16GB of RAM and at least 8GB of VRAM. The Stability AI API is also scheduled to be released soon, and the open release of version 1.0 is scheduled for mid-July.
Related Posts:
in Review, Software, Web Service, Posted by log1h_ik