giraf-imagegen

Local GPU image generation service for the GIRAF platform. Runs diffusers-based models (Flux, SDXL, etc.) behind a simple REST API.

Quick start

# Install
uv sync --extra serve

# Run (downloads model on first start)
imagegen-serve --port 8300

# Or with environment overrides
IMAGEGEN_CHECKPOINT=stabilityai/stable-diffusion-xl-base-1.0 \
IMAGEGEN_DEFAULT_STEPS=20 \
imagegen-serve

API

POST /v1/image/generate

Generate an image from a text prompt.

{
  "prompt": "A simple pictogram of a child eating lunch",
  "width": 512,
  "height": 512,
  "steps": 4,
  "guidance_scale": 0.0,
  "seed": 42,
  "format": "png"
}

Returns the image binary with appropriate Content-Type.

GET /v1/models

List loaded models and device info.

GET /health

Health check with model and device status.

Configuration

All via environment variables:

Variable	Default	Description
IMAGEGEN_CHECKPOINT	black-forest-labs/FLUX.1-schnell	HuggingFace model ID
IMAGEGEN_GPU_MEM	0.9	GPU memory fraction to use
IMAGEGEN_DEFAULT_STEPS	4	Default inference steps
IMAGEGEN_DEFAULT_WIDTH	512	Default image width
IMAGEGEN_DEFAULT_HEIGHT	512	Default image height
IMAGEGEN_MAX_WIDTH	1024	Maximum allowed width
IMAGEGEN_MAX_HEIGHT	1024	Maximum allowed height
IMAGEGEN_DTYPE	bfloat16	Torch dtype (float32, float16, bfloat16)

Docker

docker build -t giraf-imagegen .
docker run --gpus all -p 8300:8300 giraf-imagegen

Swap the FROM line in the Dockerfile for your GPU vendor (NVIDIA/AMD/Intel).

Integration with giraf-ai

Set in giraf-ai's environment:

IMAGE_PROVIDER=imagegen
IMAGEGEN_BASE_URL=http://localhost:8300