interface / externals / image_generation / img_gen_server.py

Stable Diffusion XL Lightning server. Runs on a Mac (MPS) or Jetson (CUDA) box.

Endpoints

Method · Path	Behaviour
POST /generate · POST / · POST /v1/images/generations	Body: `{"prompt", "negative_prompt", "steps"}`. Steps clamped 3-8. Guidance 1.5. 832×1216 output. Returns `{"time": float, "data": [{"b64_json": ...}]}`.

Model & scheduler

StableDiffusionXLPipeline from juggernaut_xl_lightning.safetensors.
EulerDiscreteScheduler with trailing timestep — required for Lightning models.
torch.float16 for memory.

Lifespan

asynccontextmanager loads the pipeline on startup, enables CPU offload + VAE slicing, runs a warmup pass. Garbage-collects + clears MPS/CUDA cache between generations. Sets PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 for Mac.

Dependencies

diffusers (StableDiffusionXLPipeline, EulerDiscreteScheduler)
torch (float16, mps.empty_cache)
PIL