🖼️ Text-to-Image Generator (Stable Diffusion 1.5)

A fully open-source text-to-image generator built using Stable Diffusion, PyTorch, and a clean Streamlit UI.
Supports CPU and GPU, prompt engineering, negative prompts, style presets, and metadata storage.

---

🔥 Features

🧠 AI Model

• Stable Diffusion 1.5 (open-source)
• Diffusers pipeline (HuggingFace)
• Works on CPU fallback and GPU if available

🎨 Generation

• Text prompts
• Negative prompts
• Style presets (Photorealistic, Cartoon, Cyberpunk, Van Gogh)
• Multiple image generation
• Adjustable steps, guidance scale, resolution
• Metadata saved automatically

🌐 Web UI

• Built with Streamlit
• Easy-to-use interface
• Download generated images
• Simple sliders + dropdown controls

---

📁 Folder Structure

text-to-image-generator/ │── app.py # Streamlit UI │── pipeline.py # Backend: Stable Diffusion generation │── requirements.txt # Dependencies │── README.md │ │── utils/ │ └── prompt_engineering.md # Prompt engineering guide │ │── outputs/ │ ├── images/ # Generated images │ └── metadata/ # Metadata for each image │ │── models/ # (Optional) Model storage │ └── assets/ └── banner.png # Project banner

---

🧠 Research Topics (Required by Project)

✔ Generative Adversarial Networks (GANs)

• Generator vs discriminator
• Why diffusion models outperform GANs for image generation

✔ Diffusion Models (Core of Stable Diffusion)

• Forward noise process
• Reverse denoising
• Latent Diffusion Models (LDM)
• U-Net architecture
• CLIP text encoder

✔ Prompt Engineering

• Positive and negative prompts
• Style conditioning
• Quality boost keywords
• Avoiding artifacts

---

🚀 Installation & Setup

1️⃣ Clone the repository

git clone https://github.com/Hello-H16/text-to-image-generator.git
cd text-to-image-generator
2️⃣ Create Virtual Environment
python -m venv venv
venv\Scripts\activate    # Windows
# OR
source venv/bin/activate # Mac/Linux

3️⃣ Install Dependencies
pip install -r requirements.txt

⚡ Hardware Support
🔥 GPU (Preferred)

If using NVIDIA GPU, install CUDA-enabled PyTorch:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118


Benefits:

10× faster

FP16 inference

Larger image sizes

🧠 CPU (Fallback)

Stable Diffusion runs fully on CPU:

384×384 image → ~5–12 minutes

512×512 → slower

Recommended settings:

Steps: 20–25

Size: 384×384

Guidance: 6–8

No GPU required.

🖥️ Running the App
streamlit run app.py


The app will open in your browser at:

http://localhost:8501

🎨 Sample Outputs

(Add your sample images after first generation)

<p align="center">
  <img src="samples/sample1.png" width="45%">
  <img src="samples/sample2.png" width="45%">
</p>

📦 Metadata Example

Each generated image saves metadata:

{
  "prompt": "a futuristic city at sunset",
  "negative_prompt": "blurry",
  "steps": 25,
  "guidance_scale": 7.5,
  "height": 512,
  "width": 512,
  "timestamp": "2025-03-01_134522",
  "file": "outputs/images/20250301_134522.png"
}

🧪 Prompt Engineering Guide

See:
👉 utils/prompt_engineering.md

🚧 Limitations

CPU generation is slow

Requires internet for model download (first time)

4GB+ model size

More GPU = better performance

🔮 Future Improvements

ControlNet support

Real-ESRGAN upscaling

LoRA fine-tuning

Gallery view in UI

Style mixing slider

📝 License

This project uses Stable Diffusion under the CreativeML OpenRAIL-M License.

<p align="center"> Made with ❤️ by <b>Hemant Shetty</b> </p> ```