---

⚒️ Local GenAI Forge

⚒️ Local GenAI Forge

A local, privacy-focused tool for prototyping game assets on consumer hardware.

❓ Why did I build this?

I started this project primarily to learn and explore how Generative AI pipelines work under the hood.

While experimenting, I realized that running these models locally (instead of relying on cloud APIs like Midjourney) naturally solves some critical problems in game development:

1. Privacy: Keeps game concepts and IP safe on your own machine.
2. Control: Allows for consistent results (fixed seeds) which is hard to get with cloud tools.
3. Cost: Eliminates subscription fees by utilizing consumer hardware.

This project is my implementation of that "local & private" workflow.

---

⚙️ Technical Approach

The main engineering challenge was running high-fidelity models (like DreamShaper v8) on a standard gaming laptop with limited memory (The "4GB VRAM" Challenge).

To make this work on an RTX 3050, I implemented specific optimization techniques using Python and the HuggingFace diffusers library:

• Memory Optimization (CPU Offloading): The pipeline dynamically moves model components between RAM and VRAM during inference, preventing "Out of Memory" crashes on 4GB cards.
• Precision Management:
  • Generation: Runs in Float16 for speed.
  • Decoding (VAE): Forced to Float32 with a custom VAE (mse-840000) to fix the "muddy/blurry" artifacts common in low-precision setups.
• Reproducibility System: Every generated image is automatically paired with a .txt log file containing the exact Seed, Prompt, and Settings. This ensures that any asset can be recreated perfectly by the art team later.

---

🔌 Modularity (How to improve quality)

This tool is currently configured for performance on low-end hardware. However, it is built to be modular.

You are not locked into the default model. If you have better hardware or want a different art style (e.g., Realistic, Anime, Pixel Art), you can simply change the base_model line in config.yaml.

# Example: Switching to a different model in config.yaml
model:
  base_model: "SG161222/Realistic_Vision_V5.1_noVAE"

---

🚀 Installation

I utilize uv, a Rust-based package manager, to handle dependency resolution 10-100x faster than standard pip.

Option A: The "One-Click" Setup (Windows)

I wrote a batch script to automate the entire environment setup, including the tricky CUDA/Torch bindings.

1. Clone this repository.
2. Double-click setup_windows.bat.
3. Wait for the installation to complete.
4. Double-click run_app.bat to launch the dashboard.

Option B: Manual Setup (Terminal)

# 1. Install uv (if not installed)
pip install uv

# 2. Create venv
uv venv
# Activate it: .venv\Scripts\activate (Win) or source .venv/bin/activate (Linux/Mac)

# 3. Install PyTorch (CUDA 11.8 Build)
uv pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# 4. Install Dependencies
uv pip install -r requirements.txt

# 5. Run
streamlit run ui/app.py

---

🖼️ Gallery & Benchmarks

Hardware: NVIDIA RTX 3050 (4GB VRAM) Inference Time: ~12-14 seconds per image (512x512, 40 steps)

Asset Type	Output Sample	Log Data	Raw Prompt (LLM Off)
Magma Golem		Seed: 3993855977 Step: 40 VRAM: 1.75GB	a tiny cute magma golem, made of dripping lava and obsidian rocks, glowing core, bright orange and black, angry expression, isometric view, 3d render, blender style, volumetric lighting, high contrast, masterpiece
Crystal Potion		Seed: 1708371718 Step: 40 VRAM: 1.75GB	an isometric magic potion bottle, ornate golden stopper, swirling galaxy liquid inside with glitter stars, thick ancient glass texture, caustics lighting, magical aura, dark tabletop background, crisp details, 3d render

---

🛠️ Tech Stack

• Core: Python 3.10+, PyTorch (CUDA)
• Inference: HuggingFace Diffusers, Transformers
• Optimization: Accelerate, SafeTensors
• UI: Streamlit
• LLM Integration: OpenAI API (Optional, for prompt enrichment)

---