Forma-1
Forma-1 is a diffusion model trained on 36,536 mobile UI screenshots from the
RICO dataset. Give it random noise and it will denoise it into something that
looks like a mobile app screen.
It's the model behind DiffuseUI — a project I'm building to explore generative AI applied to interface design.
Details
| Architecture | U-Net |
| Framework | Tensorflow |
| Image Size | 64x64 |
| Timesteps | 1000 |
| Noise Schedule | Linear |
| Epochs | 200 |
| Batch Size | 64 |
| Learning Rate | 1e-4 |
| Loss | MSE |
Training Data
Trained on the RICO dataset — 36,536 UI screenshots across 27 app
categories. Images were resized to 64x64 and normalized to [-1, 1] before
training.
How It Works
Standard DDPM setup. Forward process adds Gaussian noise to real UI screenshots across 1000 steps until they're pure static. The U-Net learns to predict that noise at each step. At generation time you start from pure static and denoise 1000 times — a new UI screen comes out the other end.
Limitations
- 64x64 resolution — outputs are small
- Unconditional — no control over what category of UI gets generated
- Android only — trained exclusively on Android screenshots
- 200 epochs on this dataset size produces recognizable but rough outputs
About
Built by Ricardo Flores as part of DiffuseUI.