dlp-nppe3 -- NAFNet U-Net (4x SR + denoising)

U-Net encoder/decoder built from NAFNet blocks (paper-faithful LayerNorm2d + linear channel attention), trained with PyTorch Lightning + EMA. Input: 256x160 noisy LR RGB. Output: 1024x640 clean HR RGB.

• Config: dim=48, enc_blocks=[2, 2, 4], mid_blocks=6, dec_blocks=[2, 2, 2]
• Best val_psnr (grader-exact metric): 39.137

Single checkpoint (resume-epoch=135-val_psnr=39.137.ckpt) -- the best result from a warm-restart resume run.

Load (via PyTorch Lightning)

from huggingface_hub import hf_hub_download
import sys, os

model_py_path = hf_hub_download("vedantmgoyal/dlp-nppe3", "model.py")
ckpt_path = hf_hub_download("vedantmgoyal/dlp-nppe3", "resume-epoch=135-val_psnr=39.137.ckpt")
sys.path.insert(0, os.path.dirname(model_py_path))
from model import SRLitModule

model = SRLitModule.load_from_checkpoint(ckpt_path, weights_only=False)
model.eval()

# inference
sr_img = model(lr_img_tensor).clamp(0, 1)