// the find
lllyasviel/Fooocus
Focus on prompting and generating
Fooocus is a local SDXL image generator that strips away the knob-twiddling of ComfyUI and A1111, aiming for a Midjourney-like experience where you write a prompt and get a good result without touching samplers or CFG scales. It bundles a custom k-diffusion pipeline with several non-obvious sampling improvements on top of SDXL. It is now in maintenance mode — bug fixes only, no Flux, no SD3.
The Windows installer genuinely delivers on the 'under 3 clicks' promise: unzip, run bat, models auto-download. The native refiner swap inside a single k-sampler is a real technical improvement over A1111's approach, which wastes momentum by running two independent samplers. The offline GPT-2 prompt expansion is practical — it adds cinematic language to short prompts without a round-trip to an LLM API. VRAM management is well-tuned: 4GB cards actually work, not just on paper.
The project is architecturally frozen on SDXL while the field has moved to Flux, SD3, and beyond — the maintainer explicitly says no migration is planned. The codebase is a vendored fork of ComfyUI internals in the `ldm_patched/` directory, so you're carrying a copy of someone else's code with no upstream path for fixes. The `--listen` flag exposes the UI with zero authentication by default, which is a footgun for anyone running it on a non-loopback interface. The Gradio dependency means the UI has inherent latency and is awkward to embed or automate compared to a REST API.