This project provides a self-hosted, private AI photo stylizer for users to transform images with various artistic styles, including Ghibli-inspired aesthetics. It targets users prioritizing privacy and local processing, offering a high-performance solution without cloud reliance.

How It Works

The application leverages a modern web UI built with a single-page application architecture, featuring custom style profiles, undo/redo functionality, and advanced controls. It utilizes the black-forest-labs/FLUX.1-Kontext-dev model augmented with DFloat11 quantization, which significantly reduces VRAM requirements by approximately 30% without compromising quality. This allows the full, lossless model to run on high-end consumer GPUs. The backend is a streamlined Flask server with an in-memory, thread-safe queue for asynchronous job processing, eliminating the need for external queueing systems like Redis or Celery.

Quick Start & Requirements

• Install/Run:

  git clone https://github.com/TheAhmadOsman/4o-ghibli-at-home.git
  cd 4o-ghibli-at-home
  uv venv .venv --python 3.12
  source .venv/bin/activate
  uv sync
  cp .env_template .env
  python3.12 app.py
  

• Prerequisites: Python 3.11+, uv, NVIDIA GPU with ~21GB VRAM (for DFloat11), modern web browser. Hugging Face token required for gated models.
• Setup Time: Minimal, assuming prerequisites are met.
• Links: Project Repository

Highlighted Details

• Offers dozens of built-in style profiles and custom profile management.
• Efficient VRAM usage via DFloat11 quantization, enabling consumer GPU use.
• Local storage for images and jobs with automatic cleanup.
• Simplified architecture with an in-memory task queue.

Maintenance & Community

• Primarily maintained by TheAhmadOsman.
• Support via GitHub issues for bugs, help, and feature requests.

Licensing & Compatibility

• Licensed under GNU Affero General Public License v3.0 (AGPLv3).
• Non-Commercial Use Only: Commercial use requires an explicit written license from the author.
• AGPLv3 requires derivative works deployed publicly to also be open-sourced under AGPLv3.

Limitations & Caveats

• Currently Linux-only; Windows support is planned.
• Requires significant VRAM (~21GB) for the current implementation, though future GGUF and advanced quantization support aims to lower this.
• AGPLv3 license restricts commercial use and requires derived works to be open-sourced.