1580158
/
stable-diffusion-webui-eyemask
mirror of https://github.com/ilian6806/stable-diffusion-webui-eyemask.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
stable-diffusion-webui-eyemask/CLAUDE.md

3.4 KiB
Raw Permalink Blame History

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

stable-diffusion-webui-eyemask is a Python extension for AUTOMATIC1111's Stable Diffusion Web UI. It creates masks for eyes, faces, and bodies using dlib facial landmarks or MiDaS depth estimation, then redraws those masked areas using configurable prompts and parameters.

Architecture

Entry Points (scripts/)

Three parallel entry points provide different execution modes:

  • em_script.py - Standard mode: EyeMaskScript extends scripts.Script, executes via run() method
  • em_script_embedded.py - Embedded mode: EyeMaskEmbeddedScript hooks into the generation pipeline via process()/postprocess() instead of running separately. Toggled via em_show_embedded_version setting
  • em_api.py - REST API at /sdapi/v1/eyemask/v1/* for programmatic access

Core Logic (scripts/eyemask/)

  • script.py - EyeMasksCore.execute() orchestrates the main flow: generate/receive image → generate mask → inpaint with mask prompt → save with metadata
  • script_embedded.py - EyeMasksEmbeddedCore extends EyeMasksCore, applies mask redraw as a post-processing refinement step
  • mask_generator.py - Mask detection algorithms:
    • get_eyes_mask_dlib() - Uses dlib landmarks 36-48 for eye regions
    • get_face_mask_dlib() - Uses dlib landmarks 0-27 for face outline
    • get_face_mask_depth() / get_body_mask_depth() - MiDaS depth estimation
  • ui.py - EyeMaskUI renders Gradio interface, supports both standard and embedded modes
  • wildcards.py - WildcardsGenerator handles __wildcard_name__ pattern substitution from files in /wildcards/
  • state.py - SharedSettingsContext manages model switching with automatic restoration

Modules (scripts/eyemask/modules/)

  • depthmap.py - SimpleDepthMapGenerator wraps MiDaS models (DPT Large, MiDaS v21/v21_small) with lazy loading and GPU/CPU selection

API (scripts/eyemask/api/)

  • api.py - FastAPI endpoints: /mask_list, /generate_mask, /static/{path}, /config.json
  • models.py - Pydantic request/response models
  • utils.py - Base64 image encoding/decoding

Key Patterns

Dual-Mode Execution: Standard mode runs as a separate script; embedded mode hooks into the generation pipeline for seamless integration.

Mask Types as List Indices: Mask types stored in a list with constants for easy enable/disable.

Context Manager for Model Switching: SharedSettingsContext preserves and restores checkpoint/VAE settings when temporarily switching models.

Wildcard System: Files in /wildcards/ directory; _each suffix enables sequential (vs random) line selection.

Dynamic Parameter Tracking: EM_DYNAMIC_PARAMS list enables regex-based metadata updates in saved images.

Development

Dev Mode: Enable em_dev_mode in settings to show a reload button that reloads all dynamic modules without restarting the UI.

Settings: Registered in em_settings.py, persisted via web UI shared opts and JavaScript LocalStorage.

Contributing: Submit PRs to the develop branch.

Dependencies

  • dlib (19.24.0) - Facial landmark detection
  • cmake - Required for dlib compilation (may need manual install)
  • MiDaS - Cloned automatically on first use for depth estimation
  • Pre-trained model shape_predictor_68_face_landmarks.dat downloaded to /models/