Fooocus — Focus on Prompting and Generating, Not the Tooling

Introduction

Fooocus comes from lllyasviel — the same researcher behind ControlNet — and represents an opinionated take on Stable Diffusion: hide the 50+ sliders and toggles, opinionated defaults that produce great-looking SDXL output for almost any prompt, and zero learning curve.

With over 48,000 GitHub stars, Fooocus is the recommended path for anyone who wants to "just generate good images" without learning prompt engineering, sampler theory, or VAE configuration.

What Fooocus Does

Under the hood it's a wrapper around Stable Diffusion XL. The opinionated defaults: GPT-2 prompt expansion (turns short prompts into rich descriptions), pre-tuned sampler/scheduler combinations, automatic style mixing, FreeU enhancement, and curated style presets. The UI shows a single prompt box, an aspect ratio dropdown, and a Generate button.

Architecture Overview

[Browser UI — minimal Gradio interface]
      |
[Prompt Expansion]
  GPT-2 model rewrites short prompts
  e.g. "cat in space" -> rich detailed prompt
      |
[Style Presets]
  150+ curated styles (Cinematic, Anime, Photo)
      |
[Sampler Pipeline]
  Pre-tuned: SDXL base + refiner
  FreeU enhancement
  ADetailer for faces/hands
      |
[Image Output]

Self-Hosting & Configuration

# Advanced flags
python entry_with_update.py \
  --listen \
  --port 7865 \
  --share              # Optional: expose via gradio share link

# preset files in presets/ folder
# config.txt — JSON with custom defaults if you want to override

# Sample prompts that produce great results out of the box
"a wise old wizard, dramatic lighting, fantasy art"
"futuristic city skyline, dawn, cinematic"
"product photo of a stainless steel watch on white background"
"editorial portrait of a 30-year-old woman, fashion magazine style"

Key Features

• Single prompt field — no negative prompts, no advanced sliders required
• GPT-2 prompt expansion — turns short prompts into rich descriptions automatically
• 150+ style presets — Cinematic, Anime, Photo, 3D, Pixel Art, etc.
• SDXL + Refiner — uses the highest-quality SD pipeline by default
• Image prompts — drag in reference images for variation
• Face/hand fixers — automatic detail enhancement (ADetailer-style)
• Inpaint mode — paint over an area to regenerate it
• Tiny resource budget — runs on 8GB VRAM

Comparison with Similar Tools

FAQ

Q: Fooocus vs A1111 WebUI? A: Fooocus removes choices to produce great defaults. A1111 exposes everything for power users. If you're new to SD or just want fast great-looking output, start with Fooocus.

Q: Hardware requirements? A: 8GB VRAM minimum, 12GB+ recommended. Apple Silicon supported via MPS backend. CPU mode is impractically slow.

Q: Does it support custom models? A: Yes. Drop .safetensors checkpoints into models/checkpoints/, LoRA into models/loras/. The dropdown in the Advanced tab picks them up.

Q: Can I use it as an API? A: Yes (Gradio API). For production, prefer a dedicated SD serving stack (SDXL via Diffusers + an inference server).

Sources

• GitHub: https://github.com/lllyasviel/Fooocus
• Author: lllyasviel (also created ControlNet)
• License: GPL-3.0