patilbro/21f3001839-nppe3

NPPE-3 Denoising + 4x Super-Resolution

Fine-tuned from Real-ESRGAN's RealESRGAN_x4plus (RRDBNet, 23 blocks).

Architecture

See config.json. Rebuild with:

from basicsr.archs.rrdbnet_arch import RRDBNet
import torch, json

config = json.load(open("config.json"))
model = RRDBNet(
    num_in_ch=config["num_in_ch"], num_out_ch=config["num_out_ch"],
    num_feat=config["num_feat"], num_block=config["num_block"],
    num_grow_ch=config["num_grow_ch"], scale=config["scale"],
)
model.load_state_dict(torch.load("best_model.pth", map_location="cpu"))
model.eval()

Training

Fine-tuned 50 epochs, lr=0.0002, L1Loss.