Input	CUT	CUT + Dataset Augmentation	CUT + Cycle	CUT + Cycle (Low λ)	CUT + Cycle + MultScale Discrim

📊 FID Score Comparison (↓ Lower is Better)

Fréchet Inception Distance (FID) computed against real images from the LHQ dataset. Lower scores indicate higher perceptual similarity and realism.

Input	CUT + Dataset Augmentation	CUT + Cycle	CUT + Cycle (Low λ)
198.68	114.45	106.45	108.64

📊 Mean SSIM Score (↑ Higher is Better)

Structual Similiarity Index (SSIM) computed against subsequent frames. Higher scores indicate higher perceptual similarity inbetween frames.

Input	CUT + Dataset Augmentation	CUT + Cycle	CUT + Cycle (Low λ)
0.329	0.228	0.236	0.245

TODO

Fix temporal consistency issues:

• Color consistency loss
• Gaussian filtering in temporal domain
• Conditioned training on previous frames
• Optical flow
• Self consistency loss