Dockerfile

# ImageForge — Dockerfile (CPU-only portable image)
# ===================================================
# HONEST SCOPE — read before you build this.
#
# ImageForge is designed for Apple-Silicon (MPS) and runs natively via the
# host venv (see README "Quick start" / ./start.sh). That is the recommended
# path on a Mac and the one with the proven benchmarks.
#
# Docker on macOS has NO MPS GPU passthrough, so a container can only run the
# engine on CPU — which is correct but slow (minutes per image, not seconds).
# This image therefore exists for:
#   * portability / Linux servers / CI smoke tests,
#   * exercising the HTTP API surface without a Mac,
# NOT for fast local generation on your Mac. For that, use ./start.sh.
#
# Build:
#   docker build -t imageforge .
# Run (HTTP API on host port 8765):
#   docker run --rm -p 8765:8765 imageforge
#   # then: curl http://127.0.0.1:8765/health
#
# The heavy SD weights are downloaded from HuggingFace on first generation.
# Mount a cache to avoid re-downloading across runs:
#   docker run --rm -p 8765:8765 \
#     -v "$HOME/.cache/huggingface:/root/.cache/huggingface" \
#     -v "$(pwd)/outputs:/app/outputs" imageforge

FROM python:3.11-slim

# System libs commonly needed by Pillow / diffusers image IO.
RUN apt-get update && apt-get install -y --no-install-recommends \
        libgl1 libglib2.0-0 \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /app

# Light HTTP/MCP layer first (good docker-layer caching).
RUN pip install --no-cache-dir \
        fastapi "uvicorn[standard]" mcp pydantic pydantic-settings httpx

# Heavy SD stack — CPU torch (the default torch wheel is CPU on linux/amd64).
# Kept in a separate layer so the light layer above stays cache-stable.
RUN pip install --no-cache-dir \
        torch torchvision \
        diffusers transformers accelerate safetensors Pillow huggingface_hub

# App source (models/outputs/cache are git-ignored and created at runtime).
COPY imageforge/ ./imageforge/
COPY pyproject.toml .env.example ./

# CPU device inside the container; bind on all interfaces so -p works.
ENV DEVICE=cpu \
    HOST=0.0.0.0 \
    PORT=8765 \
    IMAGEFORGE_MODEL=sd-turbo \
    OUTPUT_DIR=/app/outputs \
    CACHE_DIR=/app/cache \
    MODELS_DIR=/app/models

EXPOSE 8765

# Container healthcheck hits /health (answers even before weights load).
HEALTHCHECK --interval=30s --timeout=5s --start-period=20s --retries=3 \
    CMD python -c "import urllib.request,sys; \
        sys.exit(0 if urllib.request.urlopen('http://127.0.0.1:8765/health',timeout=3).status==200 else 1)"

CMD ["python", "-m", "uvicorn", "imageforge.api.server:app", \
     "--host", "0.0.0.0", "--port", "8765", "--log-level", "info"]