Stability AI launches the text generation picture model SDXL 0.9, and the hardware requirements are down to the consumer GPU level.
CTOnews.com June 25, Stability AI recently announced that its text-generated picture model SDXL 0.9 has been officially launched. This version makes significant improvements in image content and composition details, and can be run on consumer-grade GPU.
It is reported that SDXL 0.9 has the highest number of parameters among all open source image models, with a basic model with 3.5 billion parameters and an additional model with 6.6 billion parameters. Around these two models, SDXL works by using the basic model to create rough details, and then using additional models to refine the generated pictures. If CTOnews.com 's friends have used Stable Diffusion, they should be able to notice this gradual work process.
According to Stability AI, the SDXL team that is the source of ▲ drawings, two CLIP models are used in SDXL0.9, including OpenCLIP vitg / 14, which is by far the largest OpenCLIP model. With this model, Stable Diffusion can create realistic images with wider depth and higher resolution.
Stability AI also said that the SDXL team will release a research blog detailing the module specification and more parameter details of SDXL 0.9, which is expected to be released in July and will be open source at GitHub.