Generative-ai

Learning generative ai using various mini projects and training on big data

🖼️ Pix2Pix Experiment Summary

This project explored the Pix2Pix model from the pytorch-CycleGAN-and-pix2pix repo to understand image-to-image translation on paired datasets.

• Dataset: Maps (translating input maps (A) to satellite images (B)).
• Results: Visualizations show Input (A), Generated (Fake B), and Ground Truth (Real B).
• Note: The model was trained for only 100 epochs, and while it achieved the mapping, longer training is needed to eliminate artifacts and improve fidelity.

Pix2Pix0 Results

Input (A)	Generated (fake_B)	Ground Truth (real_B)

🦓 CycleGAN Experiment Summary

This project utilized the CycleGAN model from the pytorch-CycleGAN-and-pix2pix repository to explore unsupervised image-to-image translation.

• Dataset: horse2zebra (translating horses to zebras (A → B) and vice-versa (B → A)).
• Training: 30 epochs were completed.
• Observation: The model successfully captured the texture/style change (e.g., stripes), but more epochs are needed to fully refine the output and improve cycle consistency.
• Key Learning: Verified the effectiveness of cycle-consistency loss in learning mappings between unpaired domains.

CycleGAN Results

Input (Real A)	Generated (Fake B)