You can deploy this model and start earning money today!

tyakovenko/
night-to-day-enhancement-model-v2

Name: night-to-day-enhancement-model-v2
Availability: InStock

image-to-image

low-light-enhancement

unet

pytorch

computer-vision

tyakovenko/night-to-day-enhancement

mit

Views

No views yet

Night-to-Day Image Enhancement — v2 (L1 + MS-SSIM)

U-Net model trained to enhance low-light and night-time images to match daylight appearance. This is the v2 checkpoint, trained with a composite L1 + MS-SSIM loss instead of MSE.

Previous checkpoints (MSE-trained): tyakovenko/night-to-day-enhancement-model

Why L1 + MS-SSIM?

MSE loss minimises expected pixel error, which causes two well-known failure modes:

Washed-out / desaturated colours — MSE averages over all plausible outputs for ambiguous inputs
Blurry texture — MSE treats every pixel independently and cannot penalise spatially incoherent predictions

The v2 loss combines:

L1 (MAE) — less susceptible to outliers than MSE, produces sharper absolute values
MS-SSIM — directly optimises luminance, contrast, and structure at multiple scales

Loss weighting (Wang et al.): 0.84 * (1 - MS-SSIM) + 0.16 * L1

Model Details

Property	Value
Architecture	U-Net, 4-level encoder-decoder
Parameters	1,811,811
Base filters	16
Input/Output	3-channel RGB, float32 in [0, 1]
Loss	L1 + MS-SSIM (Wang et al. weighting)
Initialized from	`best_extended.pt` (TA+LOL, val MSE 0.027752)
Best epoch	10

Training Data

Transient Attributes — 102 outdoor scenes
LOL Dataset — 485 indoor paired images
HF Dataset repo: tyakovenko/night-to-day-enhancement
Split: 1,515 train / 161 val (scene-level)

Training Config

Hyperparameter	Value
Crop size	176×176
Batch size	4
LR	1e-5
Optimizer	Adam
LR schedule	ReduceLROnPlateau (factor=0.5, patience=4)

Evaluation Results

Channel-wise MSE (float64, values in [0, 1]):

Metric	v1-extended (MSE)	v2 (L1+MS-SSIM)
Val MSE avg	0.027752	0.028914
Val MSE — R	0.026945	0.027554
Val MSE — G	0.025522	0.025952
Val MSE — B	0.030788	0.033236

Usage

python

import torch
from PIL import Image
from torchvision import transforms
from huggingface_hub import hf_hub_download
from model import UNet  # available in this repo

ckpt_path = hf_hub_download("tyakovenko/night-to-day-enhancement-model-v2", "best_v2.pt")
ckpt = torch.load(ckpt_path, map_location="cpu")
model = UNet(base_filters=ckpt["args"]["base_filters"])
model.load_state_dict(ckpt["model"])
model.eval()

def pad_to_multiple(t, m=16):
    _, _, h, w = t.shape
    ph = (m - h % m) % m
    pw = (m - w % m) % m
    return torch.nn.functional.pad(t, (0, pw, 0, ph), mode="reflect"), h, w

img = Image.open("night.jpg").convert("RGB")
x = transforms.ToTensor()(img).unsqueeze(0)
x_padded, orig_h, orig_w = pad_to_multiple(x)

with torch.no_grad():
    out = model(x_padded)

out = out[:, :, :orig_h, :orig_w].clamp(0, 1)
transforms.ToPILImage()(out.squeeze(0)).save("enhanced.jpg")