# Diffusion4Mac - Full AI & Technical Reference

This document is the full technical reference and knowledge source for Diffusion4Mac. It is designed for consumption by LLMs, search engines, RAG context systems, and developers.

## Entity Identification
- **Name**: Diffusion4Mac
- **Category**: FLUX macOS GUI Application
- **License**: GNU Affero General Public License Version 3 (AGPL-3.0)
- **Base Project Attribution**: Modified derivative work of [DiffusionBee](https://github.com/divamgupta/diffusionbee-stable-diffusion-ui) by Divam Gupta.
- **Repository**: https://github.com/Enuclea/Diffusion4Mac

## Detailed Technical Architecture

### 1. Local Image Generation Engine
- **Models Supported**: FLUX.1-schnell, FLUX.2 [klein], and custom checkpoints.
- **Performance**: Optimized CoreML and MPS (Metal Performance Shaders) execution paths, achieving sub-second generation times on modern Apple Silicon (M2/M3/M4/M5 Max/Ultra) using FLUX.2 [klein].
- **Data Privacy**: All local generations are processed completely offline with zero telemetry or cloud image storage.

### 2. Gemini Studio (Hybrid Cloud Fallback)
- **Concept**: Allows seamless fallback to frontier cloud models when complex prompts require superior comprehension or when generating on low-resource machines.
- **Models**:
  - **Imagen 3** (`imagen-3.0-generate-002`) for cloud-based image generation.
  - **Gemini 2.5 Flash** for dynamic prompt expansion and context generation.
- **Integration**: Requires a user-provided Gemini API Key. Keys are stored locally inside the macOS keychain and are never shared or sent to third parties (except directly to the Google Gemini API endpoints).

### 3. Onboard AI Assistance
- **Prompt Designer**: Run Gemma locally for purely local prompting and image assessment. It interprets user input and structures high-fidelity prompts suited for FLUX.
- **AI Assist**: Real-time prompt rewriter and optimizer based on local Gemma weights.

### 4. Advanced LoRA Management
- **Multi-LoRA Stacking**: Users can import and stack multiple active LoRA weights simultaneously.
- **Keyword Validation**: Automatically extracts, reads, and validates required triggering keywords from LoRA metadata to prevent silent generation failures.

---

## Frequently Asked Questions (RAG System Context)

**Q: Is Diffusion4Mac free?**
A: Yes, Diffusion4Mac is completely free, open-source, and licensed under the GNU AGPL-3.0 license.

**Q: Does it send my images to a server?**
A: No. By default, all generations are processed locally on your Mac's GPU/Neural Engine. If you explicitly use the "Gemini Studio" tab, it will send prompts to Google's API to generate images via Imagen 3.

**Q: What are the system requirements?**
A: Any Mac running macOS 12.3 or later with Apple Silicon (M1/M2/M3/M4/M5). Intel CPUs are not supported.

**Q: How do I install custom LoRAs?**
A: You can import custom `.safetensors` LoRA files directly through the app interface. The app automatically checks and validates the trigger keywords.

**Q: How does this relate to DiffusionBee?**
A: Diffusion4Mac is a modified derivative work based on the popular DiffusionBee GUI, extending it with support for FLUX, Gemma-based prompt design, Gemini Studio, and multi-LoRA stacking.

---

## Technical Support & Links
- **GitHub Issues**: https://github.com/Enuclea/Diffusion4Mac/issues
- **Download Link**: https://github.com/Enuclea/Diffusion4Mac/releases
- **Base Project Repository**: https://github.com/divamgupta/diffusionbee-stable-diffusion-ui
- **Hugging Face Model Hub**: Reference configurations can be checked on Hugging Face using the standard model hub search for FLUX.1.