AI image generators have become increasingly popular, and Stability AI has recently released a new model called Stable Diffusion XL 0.9 (SDXL 0.9), which offers more realistic images compared to its predecessors. One of the key advantages of SDXL is that it allows for shorter prompts to generate the images. Version 1.0 of SDXL is expected to be released soon.
Unlike proprietary models like DALL-E and Midjourney, which are only available as cloud services with associated costs, Stable Diffusion offers an open-source model that is freely available to everyone. Stability AI applies the Creative ML OpenRAIL-M license to the model, which allows for commercial use. This is in contrast to many other open-source licenses that restrict commercial use.
The technique used in SDXL involves training a model with a dataset consisting of visual material and labels. The model then removes noise to create a copy of the original image, known as a latent image. Researchers at the University of Berkeley improved the technique by mathematically superimposing the latent images to create new images.
SDXL adds a refinement model to Stable Diffusion, allowing for the repair or scaling of images with weak resolution and pixel errors. This refinement model can be used to generate images faster and more efficiently with higher resolution. SDXL achieves an output resolution of 1024×1024 pixels, compared to the maximum output resolutions of 768×768 or 512×512 pixels for previous models.
A user study comparing SDXL to Midjourney showed a slight preference for SDXL in terms of prompt compliance. However, both models have their own strengths and weaknesses, and the choice depends on the specific application and requirements.
When using SDXL, it’s important to note that German prompts are possible, but the results may not achieve the same quality as English prompts. The model is particularly suitable for generating photorealistic images and strong cartoon images. However, there may be issues with the generation of hands, and fonts may be generated more accurately compared to previous models.
Overall, SDXL offers improved performance and flexibility with its open-source license. It is one of the leading models in the field of AI image generation. However, it’s important to stay updated with the rapidly evolving developments in this field. SDXL 1.0 is expected to be released in July 2023.