Research map / Image generation / Editing and style transfer
SEIT: Structural Enhancement for Unsupervised Image Translation in Frequency Domain
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 SEIT, a one-sided GAN for unsupervised image translation that swaps Fourier amplitude for style (FDA module), enhances high-frequency wavelet sub-bands of translated features (WSE module), and uses multi-scale domain-specific encoders.
Problem
Unsupervised image-to-image translation struggles to transfer target-domain style while preserving the source image's structure, with GAN outputs often showing structural distortion and artifacts.
Approach
Decouple content and style in the frequency domain: replace the source Fourier amplitude with a target-conditioned stylized amplitude while keeping source phase, and enrich high-frequency wavelet sub-bands of the translated features with source high-frequency information.
Key results
- On GTA5→Cityscapes, SEIT reaches SSIM 0.93, FSIM 0.95 and FID 82.55, the best of all compared methods.
- On SYNTHIA Day→Night, SEIT achieves SSIM 0.87 and FID 46.15, versus 0.80 and 48.68 for the next-best VSAIT.
- Ablation on SYNTHIA shows the full model (multi-scale + FDA + WSE) reaches SSIM 0.87 and FID 46.15, improving over multi-scale alone (0.80 SSIM, 46.44 FID in its row).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SEIT instead of the strongest baseline | SYNTHIA Day →Night · SYNTHIA Day →Night | SSIM: 0.83 -> 0.87 | Helped |
| Use SEIT instead of the strongest baseline | SYNTHIA Day →Night · SYNTHIA Day →Night | FID: 48.68 -> 46.15 | Helped |
| Use SEIT instead of the strongest baseline | BDD Sunny →Cloudy · BDD Sunny →Cloudy | SSIM: 0.92 -> 0.94 | Helped |
| Use SEIT instead of the strongest baseline | BDD Sunny →Cloudy · BDD Sunny →Cloudy | FID: 41.69 -> 40.74 | Helped |
| Use SEIT instead of the strongest baseline | GTA5 →CityScapes · GTA5 →CityScapes | SSIM: 0.77 -> 0.93 | Helped |
| Use SEIT instead of the strongest baseline | GTA5 →CityScapes · GTA5 →CityScapes | FID: 82.77 -> 82.55 | Helped |
| Add FDA to the single-scale architecture (ablation) | SEIT single-scale · SYNTHIA Day →Night | SSIM: 0.54 -> 0.67 | Helped |
| Add WSE to the single-scale architecture (ablation) | SEIT single-scale · SYNTHIA Day →Night | SSIM: 0.54 -> 0.64 | Helped |
| Add FDA and WSE to the single-scale architecture (ablation) | SEIT single-scale · SYNTHIA Day →Night | SSIM: 0.54 -> 0.70 | Helped |
| Use multi-scale instead of single-scale architecture (ablation) | SEIT · SYNTHIA Day →Night | SSIM: 0.54 -> 0.78 | Helped |
| Add FDA to the multi-scale architecture (ablation) | SEIT multi-scale · SYNTHIA Day →Night | SSIM: 0.78 -> 0.80 | Helped |
| Add WSE to the multi-scale architecture (ablation) | SEIT multi-scale · SYNTHIA Day →Night | SSIM: 0.78 -> 0.83 | Helped |
Limitations
- Style transfer is limited to the trained source-target pairs; multi-modal extension is proposed only as future work.
- Reported results are on three translation tasks using only SSIM, FSIM and FID, with no human evaluation.
Experimental setup
Datasets: SYNTHIA GTA5 Cityscapes BDD
Benchmarks: SYNTHIA Day-to-Night BDD Sunny-to-Cloudy GTA5-to-Cityscapes
Metrics: SSIM FSIM FID
Baselines: MUNIT LPTN CUT TSIT F-Sesim VSAIT
Builds on: TSIT LPTN CycleGAN Fourier transform Haar wavelet
Closest papers by meaning
- W-EDIT: A Wavelet-Based Frequency-Aware Framework for Text-Driven Image Editing (ICLR 2026 · similarity 0.84)
- FDS: Frequency-Aware Denoising Score for Text-Guided Latent Diffusion Image Editing (CVPR 2025 · similarity 0.83)
- Disentangling Dual Image References in Frequency Aware Diffusion Models for Personalized Generation (NeurIPS 2026 · similarity 0.83)
- Misalignment-Robust Frequency Distribution Loss for Image Transformation (CVPR 2024 · similarity 0.82)
- Uncertainty-Aware Source-Free Adaptive Image Super-Resolution with Wavelet Augmentation Transformer (CVPR 2024 · similarity 0.82)
- Learning Frequency-Adapted Vision Foundation Model for Domain Generalized Semantic Segmentation (NeurIPS 2024 · similarity 0.82)
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.