README.md

metadata

title: Draw Your Floorplan - ControlNet
emoji: 🏠
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit
models:
  - Qistinasofea/controlnet-floorplan
  - stable-diffusion-v1-5/stable-diffusion-v1-5

🏠 Draw Your Floorplan - ControlNet

AI54 Final Project - Spatially Conditioned Floorplan Generation

🎨 Interactive Demo

This Space allows you to draw colored segmentation masks and generate architectural floorplans using a fine-tuned ControlNet model.

How to Use:

1. Draw colored regions on the canvas - each color represents a different room type
2. Describe your floorplan in the text box
3. Adjust settings if needed (inference steps, control strength, seed)
4. Click Generate Floorplan to see your AI-generated layout!

Suggested Colors:

• 🔴 Red - Living room / Main spaces
• 🟢 Green - Bedrooms
• 🔵 Blue - Bathrooms
• 🟡 Yellow - Kitchen
• 🟣 Purple - Dining area
• 🟠 Orange - Office / Study
• 🩵 Cyan - Utility / Storage

---

📊 Model Information

Training Details:

• Method: Full ControlNet Fine-Tuning
• Base Model: Stable Diffusion 1.5 (frozen)
• ControlNet: Segmentation variant (fully trained)
• Dataset: 11,375 orientation-normalized floorplan samples
• Parameters: 361M trainable parameters (100% of ControlNet)
• Training Steps: 10,000
• Final Loss: 0.0887
• Training Time: 3.7 hours on T4 GPU

Architecture:

The model uses a two-stage architecture:

1. Base Model (SD 1.5): Generates realistic textures and appearance (frozen weights)
2. ControlNet: Guides spatial structure based on colored segmentation input (fully fine-tuned)

This separation allows the model to:

• ✅ Preserve spatial layouts from user drawings
• ✅ Generate realistic architectural details
• ✅ Maintain consistent room boundaries
• ✅ Produce diverse outputs from the same layout

---

🔗 Links

• Trained Model: Qistinasofea/controlnet-floorplan
• Dataset: Qistinasofea/floorplan-12k-aligned
• Training Notebook: Available in model repository

---

🎓 Academic Context

This is a final project for AI54: Artificial Intelligence course focused on:

• Conditional image generation
• Spatial control in diffusion models
• ControlNet architecture and training
• Parameter-efficient fine-tuning considerations
• Real-world application development

Key Contributions:

1. Dataset Preprocessing: Orientation normalization using PCA-based rotation alignment
2. Training Strategy: Full fine-tuning justified by dataset size (11,375 samples)
3. User Interface: Visual layout-driven interaction for non-technical users

---

💻 Technical Stack

• Framework: 🤗 Diffusers
• Model: ControlNet + Stable Diffusion 1.5
• Interface: Gradio
• Deployment: HuggingFace Spaces
• Hardware: GPU-enabled (T4 or better recommended)

---

📝 Citation

If you use this model or approach in your work, please cite:

@misc{controlnet-floorplan-2024,
  author = {Qistinasofea},
  title = {ControlNet for Floorplan Generation},
  year = {2024},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/Qistinasofea/controlnet-floorplan}}
}

---

📄 License

This project is released under the MIT License. The base Stable Diffusion 1.5 model follows its original CreativeML Open RAIL-M license.

---

Built with ❤️ for AI54 Final Project