Pranavz/minor32-fog-removal

minor32 Fog Removal Models

Fine-tuned DehazeFormer-S checkpoints for the minor32 fog-removal project on RESIDE-IN.

Included checkpoints

• checkpoints/dehazeformer_s_reside_15ep.pth
• checkpoints/dehazeformer_s_reside_50ep.pth
• checkpoints/dehazeformer_s_reside_300ep.pth

Main result

The final project checkpoint is dehazeformer_s_reside_300ep.pth.

Evaluation on the RESIDE-IN test split:

• 15 epochs: PSNR 22.93, SSIM 0.8926
• 50 epochs: PSNR 25.28, SSIM 0.9111
• 300 epochs: PSNR 27.56, SSIM 0.9451

Files

• metrics/: raw eval result JSON files for the three runs
• configs/: training config snapshots
• notebook/fog_removal_project_showcase.ipynb: one-stop project notebook
• notebook_assets/: qualitative comparison images and result tables used by the notebook
• demo_inputs/: default sample hazy/reference clips used by the notebook
• demo/: sample output video and metrics from the notebook inference path

Notes

• These are project checkpoints trained from the official DehazeFormer codebase, wrapped by the minor32 training/runtime pipeline.
• Training was performed remotely on Modal GPUs.
• The local notebook/runtime can run on CPU or GPU for inference. GPU is recommended for longer or higher-resolution videos.
• The notebook is designed to auto-download missing checkpoints and bundled demo assets from this Hugging Face repo.