confyui

KπX sovereign local image generation and editing daemon over ComfyUI.

Hardware-accelerated on Intel Arc (XPU) via PyTorch XPU wheels, with portable CPU and CUDA extras.

Philosophy

• 0 Magic: The CLI talks to ComfyUI's standard REST API via structured JSON graphs.
• 0 Manual Start: The first call (generate, edit, inpaint) automatically starts ComfyUI in the background.
• 0 Wasted RAM: An idle watchdog shuts ComfyUI down after CONFYUI_IDLE_TIMEOUT seconds of inactivity (default: 30 minutes).
• 0 Hardcoding: Paths expand $HOME at runtime and are 100% overridable via environment variables.

Installation

# Clone or navigate to the repository
cd ~/Labs/KpihX-Labs/AI/confyui

# Intel Arc workstation (XPU)
make sync-xpu
# or: uv sync --extra xpu

# Generic CPU fallback
make sync-cpu

# NVIDIA CUDA host
make sync-cuda

Environment Variables

Usage

1. Text-to-Image (generate)

# Minimal
uv run confyui generate "a red panda wearing a top hat on the moon, cinematic lighting"

# Full options
uv run confyui generate "a red panda wearing a top hat on the moon" \
    --negative "blurry, low quality, artifacts" \
    --width 512 \
    --height 512 \
    --steps 20 \
    --cfg 7.0 \
    --seed 42 \
    --sampler euler \
    --scheduler normal \
    --model "Realistic_Vision_V5.1_noVAE.safetensors" \
    --out panda_moon.png

2. Image-to-Image (edit)

Modifies an existing image based on a prompt. The source image is kept and only altered proportionally to --denoise (0.0 = unchanged, 0.45-0.60 = balanced modification, 1.0 = total repaint):

uv run confyui edit panda_moon.png "make the red panda wear a golden astronaut suit" \
    --denoise 0.50 \
    --steps 20 \
    --out panda_astronaut.png

3. Inpainting (inpaint)

Repaints only the masked region (white pixels in the mask are regenerated, black pixels preserved):

uv run confyui inpaint portrait.png torso_mask.png "she wears a long red silk dress" \
    --denoise 1.0 \
    --steps 25 \
    --out portrait_red_dress.png

4. Daemon Management

# Inspect status, uptime, PID, and idle time remaining
uv run confyui status

# List installed checkpoints
uv run confyui models

# Stop the daemon immediately to free RAM and VRAM
uv run confyui stop

Development

make help     # Show available targets
make check    # Lint (ruff), format check, compile, pyright, pytest
make smoke    # Fast sanity check of CLI and imports
make demo     # Run a real 12-step test generation