Stable Diffusion is a powerful generative AI model for images, but operating it via web and local interfaces often involves a lot of tweaking and prompt engineering. Fooocus aims to change that.
Lvmin Zhang, the person behind Fooocus, describes the project as a reworking of the design of Stable Diffusion and Midjourney. From Stable Diffusion, Fooocus takes the model and the focus on offline capabilities and open source, from Midjourney the focus on ease of use: manual adjustments of values like CFG are not required, users can simply focus on the image and the prompt.
In short, Fooocus is like a free offline version of Midjourney using the latest SDXL model from Stability AI. Although Midjourney usually still gives better results in my short test, Fooocus with SDXL comes close.
Fooocus has low requirements and easy installation
Fooocus comes with a simple installation, and the number of mouse clicks between hitting “download” and generating the first image is kept to a minimum, promises Zhang.
Behind the scenes, the project has built-in and automated many optimizations and quality improvements that have to be set manually for Stable Diffusion in other interfaces. As with Midjourney, this should give good results on every attempt. If you want to do more, you can use the Advanced tab in Fooocus. Here you can set a sharpness filter or custom LoRAs, for example. You can also set the style by making a simple selection.
It requires at least an Nvidia graphics card with 4 gigabytes of RAM and 8 gigabytes of system RAM under Windows. Microsoft’s Virtual Swap needs to be enabled, but this is usually done automatically and can be turned on relatively easily. On a laptop with 16 gigabytes of system RAM and an Nvidia 3060 with 6 gigabytes of RAM, Zhang reports that image generation takes less than 1.5 seconds.
More information and the download are available on the Fooocus GitHub.