Work    Notes    Info

AI Lounge Chair 01
(Silhouette from AI)


2023
Self-initiated


Stainless steel frame with wicker seat and back.
Photography by Studio Pan, Lee Jin Woo



This project explores the use of AI in the idea sketching phase of the industrial design process, which is the moment of initial archetype creation and the source of creativity. A proprietary, small-scale AI model was developed using open-source code, with the following methodology:

1. A dataset of chair images from the Modernist era (1920-1970) was curated.
2. The AI model was trained using deep learning algorithms (DDPM).
3. Images were generated through the AI model.









Algorithm: Denoising Diffusion Probabilistic Models (DDPM)
Architecture: UNet2DModel
Training Type: Unconditional

Image Preprocessing:
- Padding
- Flip All Images
- Add Rescaled Images (x2)
- Resize to 256 x 256
- Center Crop

Hyperparameters:
- Batch Size: 2
- Epochs: 500
- Learning Rate: 1e-4 (0.0001)
- EMA(Exponential Moving Average) is used

Optimization:
- Gradient Accumulation Steps: 2
- Mixed Precision: bf16

Image Generation:
- Device: CUDA (GPU)
- Epochs: 330-490
- Number of inference steps (/image): 50
- Batch size for generation: 4-16

Final Image: Performed inference at 330 epochs
© 2025 HyungGyu Rhee