FastAPI backend for text-to-image generation powered by FLUX.2 [klein] 4B.

• 🎨 Text-to-image generation with FLUX.2 [klein] 4B
• ⚡ Sub-2-second inference (4 steps, step-distilled)
• 🌐 Built-in web demo at /
• 📊 Request queue with async processing
• 🔧 Configurable via environment variables

git clone https://github.com/jamarju/flux2-api-v3.git
cd flux2-api-v3
uv run flux2-api

Open http://localhost:8000 for the web demo.

Generate an image from a text prompt.

Request:

{
  "prompt": "A cat holding a sign that says hello world",
  "seed": 42,
  "width": 1024,
  "height": 1024,
  "num_inference_steps": 4
}

All parameters except prompt are optional.

Response: PNG binary image

Response Headers:

Example with curl:

curl -X POST http://localhost:8000/generate \
  -H "Content-Type: application/json" \
  -d '{"prompt": "A sunset over the ocean"}' \
  -o output.png

Health check endpoint.

Response:

{"status": "ok", "model_loaded": true}

Environment variables (prefix FLUX2_):

• Python ≥ 3.12
• NVIDIA GPU with ≥ 13 GB VRAM (RTX 3090/4070+)
• uv package manager
• CUDA toolkit

See docs/cloudflare-tunnel.md for instructions on exposing the service to the internet via Cloudflare Tunnel.

# Unit tests (requires GPU)
uv run pytest

# Smoke test (starts a real server)
uv run pytest -m smoke

For the full report, see benchmarks/results/REPORT.md.