night-to-day-enhancement-model

Night-to-Day Image Enhancement — Model Repo

U-Net models trained to enhance low-light and night-time images to match daylight appearance.

This repo contains two checkpoints:

---

Architecture

• Model: U-Net, 4-level encoder-decoder with skip connections
• Parameters: 1,811,811
• Base filters: 16
• Input/Output: 3-channel RGB, float32 in [0, 1]
• Output activation: Sigmoid

---

best.pt — v1 (Transient Attributes only)

Training Data

• Dataset: Transient Attributes — 102 outdoor scenes, time-of-day captures
• HF Dataset: tyakovenko/night-to-day-enhancement
• Split: 1,010 train / 166 val (scene-level)

Training Config

Results

---

best_extended.pt — v1-extended (TA + LOL fine-tune)

Fine-tuned from best.pt. Adds the LOL indoor dataset to improve generalisation to non-outdoor scenes.

Training Data

• Datasets: Transient Attributes + LOL (485 indoor paired images)
• Combined split: 1,515 train / 161 val

Training Config

Results

---

Usage

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

REPO = "tyakovenko/night-to-day-enhancement-model"

# Choose checkpoint: "best.pt" or "best_extended.pt"
ckpt_path = hf_hub_download(REPO, "best_extended.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):
    """Pad spatial dims to a multiple of m (required by 4-level U-Net)."""
    _, _, 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")

---

Limitations

• Both checkpoints use MSE loss, which can produce mildly washed-out results on very dark inputs
• best.pt trained on outdoor scenes only — may underperform on indoor images
• best_extended.pt improves indoor generalisation but the LOL weighting is not tuned; outdoor performance is largely preserved
• See tyakovenko/night-to-day-enhancement-model-v2 for a version trained with L1 + MS-SSIM loss for improved perceptual quality