Research map / Image generation / Editing and style transfer
DNIT: Enhancing Day-Night Image-to-Image Translation through Fine-Grained Feature Handling (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DNIT, a day-night image-to-image translation method adding a nighttime preprocessing module (SCI brightening plus DCP defogging) and an edge-fusion loss (Sobel and Canny) on top of SPN2D-GAN.
Problem
Image-to-image translation between day and night domains performs poorly because nighttime images have low visibility and structural details are easily lost.
Approach
Preprocess nighttime inputs with self-calibrated illumination (SCI) brightening followed by dark channel prior (DCP) denoising/defogging before the generator, and add an edge loss that combines Sobel and Canny edge maps of real and generated images as extra generator constraints.
Key results
- DNIT lowers FID from 65.54 (SPN2D-GAN) to 64.72, KID from 2.46 to 2.23, and BRISQUE from 20.36 to 19.04 on ACDC.
- NIQE is slightly worse (2.65 vs 2.58 for SPN2D-GAN).
- In a user study with 20 evaluators ranking 70 images, DNIT was ranked top-1 in 51% of cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the NIP and EFD modules to SPN2D-GAN | DNIT · ACDC | FID: 65.54 -> 64.72 | Helped |
| Add the NIP and EFD modules to SPN2D-GAN | DNIT · ACDC | KID: 2.46 -> 2.23 | Helped |
| Add the NIP and EFD modules to SPN2D-GAN | DNIT · ACDC | BRISQUE: 20.36 -> 19.04 | Helped |
| Add the NIP and EFD modules to SPN2D-GAN | DNIT · ACDC | NIQE: 2.58 -> 2.65 | Hurt |
Surprising findings
- DNIT has a slightly higher NIQE (2.65) than the SPN2D-GAN baseline (2.58), which the authors attribute to lower natural scene statistics quality.
Limitations
- Canny threshold selection is manual/fixed, limiting generalization across diverse scenes.
- NIQE is slightly worse than the baseline.
- Evaluated only on the ACDC dataset in a short abstract-length study.
Experimental setup
Datasets: ACDC
Benchmarks: ACDC
Metrics: FID KID BRISQUE NIQE user study ranking
Baselines: CycleGAN CUT NICE-GAN DCLGAN ToDayGAN SPN2D-GAN
Builds on: SPN2D-GAN CycleGAN SCI (self-calibrated illumination) dark channel prior
Closest papers by meaning
- Rethinking Nighttime Image Deraining via Learnable Color Space Transformation (NeurIPS 2025 · similarity 0.84)
- NightAdapter: Learning a Frequency Adapter for Generalizable Night-time Scene Segmentation (CVPR 2025 · similarity 0.84)
- Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow (ICLR 2024 · similarity 0.83)
- VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook (AAAI 2024 · similarity 0.83)
- NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction (AAAI 2024 · similarity 0.83)
- Unveiling Details in the Dark: Simultaneous Brightening and Zooming for Low-Light Image Enhancement (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.