Image Generation API

Production-ready FastAPI image generation service powered by diffusers with 4-step distilled models.

Features

• 4-step fast generation with distilled/Lightning models
• Qwen-Image-2512 FP8 Lightning - High quality, 4 steps
• FLUX.2 Klein 4B - Fast distilled generation, 4 steps
• OpenAI-compatible API format
• Automatic model caching via HuggingFace Hub
• VRAM management with auto-clearing between generations

Available Models

Model Details

qwen-2512-fp8-4step:

• Source: lightx2v/Qwen-Image-2512-Lightning
• File: qwen_image_2512_fp8_e4m3fn_scaled_4steps_v1.0.safetensors (20.5 GB)
• Format: FP8 (e4m3fn) quantized with 4-step Lightning distillation baked in
• Max Resolution: 1664x1664

flux-klein-4b:

• Source: black-forest-labs/FLUX.2-klein-4B
• Format: Native 4-step distilled model
• Max Resolution: 1280x1280

Supported Resolutions

Qwen Lightning:

FLUX Klein:

Quick Start

Docker Deployment (Recommended)

# Clone repository
git clone https://github.com/gordo-v1su4/image-gen-endpoint.git
cd image-gen-endpoint

# Build and run with GPU support
docker compose up --build -d

# View logs
docker compose logs -f

Models download automatically from HuggingFace on first use (~30GB total).

API Usage

Generate with Qwen Lightning (default):

curl -X POST https://your-server.com/v1/images/create \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A majestic lion in the African savanna at golden hour",
    "width": 1328,
    "height": 1328,
    "steps": 4,
    "guidance_scale": 1.0
  }'

Generate with FLUX Klein:

curl -X POST https://your-server.com/v1/images/create \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A cyberpunk cityscape at night",
    "model": "flux-klein-4b",
    "width": 1024,
    "height": 1024,
    "steps": 4
  }'

Response Format (OpenAI-compatible):

{
  "data": [{"b64_json": "...base64 encoded image..."}],
  "created": 1706500000,
  "model": "qwen-2512-fp8-4step"
}

Endpoints

Requirements

• GPU: NVIDIA RTX 4090 (24GB VRAM) or equivalent with 20GB+ VRAM
• CUDA: 12.0+
• Docker: With NVIDIA Container Toolkit configured
• Storage: ~40GB for cached models

Server Setup

# Install NVIDIA Container Toolkit
sudo apt-get install nvidia-container-toolkit

# Configure Docker runtime
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

# Verify GPU access
docker run --rm --gpus all nvidia/cuda:12.4.1-base-ubuntu22.04 nvidia-smi

Environment Variables

Architecture

┌─────────────────────────────────────┐
│         FastAPI Application         │
├─────────────────────────────────────┤
│         Model Manager               │
│  ┌─────────────┐ ┌───────────────┐  │
│  │Qwen Lightning│ │ FLUX Klein   │  │
│  │  (diffusers) │ │ (diffusers)  │  │
│  └─────────────┘ └───────────────┘  │
├─────────────────────────────────────┤
│         HuggingFace Hub             │
│      (automatic model caching)      │
└─────────────────────────────────────┘

Model Sources

• Qwen FP8 Lightning: lightx2v/Qwen-Image-2512-Lightning
• FLUX Klein 4B: black-forest-labs/FLUX.2-klein-4B

License

Apache 2.0 - See LICENSE for details.