Cyclone: Diffusion Model for Cycle-Consistent Weather Editing from Unpaired Driving Data
Merged summary
TL;DR — Cyclone is a latent diffusion framework for editing weather conditions in driving imagery without paired training data, aiming to improve autonomous-driving perception robustness across adverse conditions.
- Uses latent diffusion with cycle-consistent constraints plus knowledge from image-text models to synthesize multiple weather conditions across diverse scenes, removing the need for paired data.
- Claims more realistic, structure-preserving outputs than existing baselines (vs. synthetic augmentation or physics-based, task-specific methods).
- Reports consistent gains on several downstream driving perception tasks.
- Can be distilled into a video diffusion model for temporally consistent weather editing.
Sources (1)
Cyclone: Diffusion Model for Cycle-Consistent Weather Editing from Unpaired Driving Data
TL;DR — Cyclone is a latent diffusion framework for editing weather conditions in driving imagery without paired training data, aiming to improve autonomous-driving perception robustness across adverse conditions.
- Uses latent diffusion with cycle-consistent constraints plus knowledge from image-text models to synthesize multiple weather conditions across diverse scenes, removing the need for paired data.
- Claims more realistic, structure-preserving outputs than existing baselines (vs. synthetic augmentation or physics-based, task-specific methods).
- Reports consistent gains on several downstream driving perception tasks.
- Can be distilled into a video diffusion model for temporally consistent weather editing.