thethinkmachine/nafnet-ll-ssr-x4

Low-Light Denoising + 4× Super-Resolution (NAFNet)

A joint denoising and 4× super-resolution model for noisy low-light images, mapping 256×160 LR inputs to 1024×640 HR outputs.

Recovered Degradation Model

The image degradation process recovered from the training data is:

LR = round(clip(box4(HR) + N(0, sqrt(2.514 × I + 2.037))))

where box4 denotes 4× box downsampling and the noise follows an intensity-dependent Gaussian model.

Load the Model

from huggingface_hub import hf_hub_download
import torch

ckpt_path = hf_hub_download(
    repo_id="piushdasss/nppe-lowlight-sr-nafnet",
    filename="model.pt",
)

ckpt = torch.load(
    ckpt_path,
    map_location="cpu",
)

Results

Metric	Result
Test-like holdout PSNR	37.131 dB
Official validation PSNR	39.3 dB
Model parameters	2.30M

Low-Light Denoising + 4× Super-Resolution (NAFNet)

A joint denoising and 4× super-resolution model for noisy low-light images, mapping 256×160 LR inputs to 1024×640 HR outputs.

Recovered Degradation Model

The image degradation process recovered from the training data is:

LR = round(clip(box4(HR) + N(0, sqrt(2.514 × I + 2.037))))

where box4 denotes 4× box downsampling and the noise follows an intensity-dependent Gaussian model.

Load the Model

from huggingface_hub import hf_hub_download
import torch

ckpt_path = hf_hub_download(
    repo_id="piushdasss/nppe-lowlight-sr-nafnet",
    filename="model.pt",
)

ckpt = torch.load(
    ckpt_path,
    map_location="cpu",
)

Results

PSNR Set	Result
Holdout	37.131 dB
Validation set	39.3 dB
Model parameters	2.30M