Fooosti - Fast Fooocus API

Fooosti is a self-hosted fork of Fooocus, tuned to run as a lightweight background service in Docker. It keeps the familiar Fooocus image quality and UI, and adds a clean HTTP API on top.

Why a fork?

• One process owns everything. A single fooosti daemon launches the WebUI, the API and the generation worker (a subprocess) — no fragile chain of scripts to keep alive.
• Runs idle for almost nothing. Generation happens in a worker subprocess that is spawned per task and torn down when done, so the main process sits nearly idle in between — no tens of gigabytes pinned in VRAM while you do nothing else on the GPU. The worker lifetime is controlled by the WORKER_KEEPALIVE_MINUTES env var (0 = unload the worker after every task).
• Lazy clients by default. The WebUI and API can live inside the daemon (always-alive, ~176 MiB idle) or be spawned on demand (WEBUI_KEEPALIVE_MINUTES / API_KEEPALIVE_MINUTES, 0 = lazy): the daemon holds their port, starts the client on the first request and takes the port back after it exits on idle — leaving the container at 3-11 MiB (depending on swap) until something touches it.
• A real API. FastAPI server on port 8890 with A1111-compatible /sdapi/v1/* endpoints — so it plugs straight into Open WebUI (or any other A1111 client) as a drop-in backend. API requests are queued and processed one at a time by the single worker. See API.md for endpoints, request templates and auth.
• Local prompt translation & enhancement. Non-English prompts are rewritten into detailed English SD prompts by a local LLM — Qwen2.5-1.5B-Instruct-abliterated, downloaded from HuggingFace. Works on top of the stock Fooocus V2 expansion. Fully configurable via config.txt: enable_prompt_translator disables it entirely, prompt_translator_enhance_english also enhances English prompts.
• Redesigned UI. Custom CSS and JS on top of the Gradio interface, UI settings persisted across restarts.

Screenshots

Fullscreen	Medium screen	Mobile screen

Quick start (Docker)

Requirements: Nvidia GPU with proprietary drivers + Docker with the NVIDIA Container Toolkit.

mkdir -p ~/Fooosti
cd ~/Fooosti
docker compose up -d --build

• Web UI (Gradio): http://localhost:7865
• API (FastAPI): http://localhost:8890 (interactive docs at /docs)
• Data lives on the persistent fooocus-data Docker volume (mounted at /content/data) — checkpoints, LoRAs, outputs, config and the translator model survive container rebuilds.

Drop .safetensors checkpoints into the volume with docker compose cp model.safetensors app:/content/data/models/checkpoints/, press Refresh All Files in the UI (Models tab, bottom of the right column), and you're done.

To run only one part of the stack, set CMDARGS=--serves webui or CMDARGS=--serves api in docker-compose.yml (--serves both is the default).

Auth

The API can be locked with the FOOOSTI_API_TOKEN env var in docker-compose.yml (optional; empty by default). When set, every /sdapi/v1/* request must send the token in the x-api-token header. Empty token = unauthenticated. Details: API.md.

More docs

• API.md — endpoints, request templates, auth.
• docker.md — running with plain Docker, Podman.
• Upstream Fooocus README — original docs, install guides, FAQ and troubleshooting (swap, VRAM, CUDA errors).

License

GPL-3.0 — inherited from the upstream lllyasviel/Fooocus. See LICENSE.