NPPE-3: Low-Light Denoising and 4x Super-Resolution

Model Architecture

• Backbone Architecture: Swin2SR (Swin Transformer V2 for Image Super-Resolution)
• Base Pretrained Checkpoint: caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr
• Scale Factor: 4x
• Input Window Size: 8
• Embedding Dimension: 180
• Depths: [6, 6, 6, 6, 6, 6]
• Number of Attention Heads: [6, 6, 6, 6, 6, 6]
• Training Loss: Edge-Preserving Charbonnier Loss
• Best Validation PSNR: 39.4857 dB

Intended Use & Inference

import torch
from transformers import Swin2SRForImageSuperResolution
from PIL import Image
import torchvision.transforms.functional as TF

model = Swin2SRForImageSuperResolution.from_pretrained("madhav512/nppe3-swin2sr-lowlight-denoise")
model.eval()

img = Image.open("noisy_lowlight.png").convert("RGB")
pixel_values = TF.to_tensor(img).unsqueeze(0)

with torch.no_grad():
    output = model(pixel_values=pixel_values).reconstruction
output_img = TF.to_pil_image(torch.clamp(output.squeeze(0), 0.0, 1.0))
output_img.save("enhanced_clean_4x.png")