Fooocus

Fooocus wraps Stable Diffusion XL in a Gradio interface that deliberately hides nearly every knob, leaving the user to write prompts and, optionally, supply reference images. Behind that sparse screen it applies its own sampling adjustments, a prompt expansion step that runs a small GPT-2 model locally, a custom inpainting model, and its own image-prompt method in place of standard IP-Adapter, along with upscale and variation buttons, outpainting in four directions, face swap and an image describer. Presets choose the checkpoint, with general, anime and realistic launchers, and SDXL checkpoints from Civitai can be dropped in. Windows users unpack a 7z archive and run a batch file; Linux, Docker, Apple Silicon and AMD routes through DirectML or ROCm are documented, the AMD path as beta, and the stated floor is an NVIDIA card with 4 GB of VRAM alongside 8 GB of system RAM. The code is GPL-3.0. The project now sits in a bug-fix-only long-term support mode with no plans to adopt Flux or other newer architectures, so it fits beginners who want polished SDXL results today rather than users chasing the latest models.