Research map / Image generation / Image restoration
STAR: A Structure-Aware Lightweight Transformer for Real-Time Image Enhancement
Zhaoyang Zhang, Yitong Jiang, Jun Jiang, Xiaogang Wang, Ping Luo, Jinwei Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace CNN backbone with STAR | DCE-Net · FiveK low-light | PSNR (dB): 22.2 -> 24.5 | Improved |
| replace CNN backbone with STAR | DCE-Net · FiveK low-light | parameters and FLOPS: 79.4K parameters, 0.51G FLOPS -> 20.1K parameters, 0.04G FLOPS | Improved |
| use Mean head tokenization instead of Linear head (ablation) | STAR-DCE · FiveK low-light | PSNR (dB): 24.0 -> 24.5 | Improved |
| use Mean head tokenization instead of Conv head (ablation) | STAR-DCE · FiveK low-light | PSNR (dB): 24.5 -> 24.5 | No clear effect |
| add the two-branch long-short range design (ablation) | STAR-DCE (Mean head) · FiveK low-light | parameters (K): 23.3 -> 20.1 | Improved |
| add the two-branch long-short range design (ablation) | STAR-DCE (Mean head) · FiveK low-light | PSNR (dB): 24.5 -> 24.5 | No clear effect |
| reduce CNN depth from 7 to 3 layers (ablation) | DCE-Net · FiveK low-light | PSNR (dB): 22.7 -> 12.3 | Hurt |
| reduce CNN width from 32 to 16 (ablation) | DCE-NetD · FiveK low-light | PSNR (dB): 22.2 -> 21.3 | Hurt |
| reduce STAR width from 32 to 16 (ablation) | STAR-DCE · FiveK low-light | PSNR (dB): 24.5 -> 23.0 | Hurt |
| replace CNN encoder with STAR | WB editing · Cube+ | MAE: 3.45 -> 3.31 | Improved |
| replace CNN encoder with STAR | WB editing · Rendered WB Set 1-Test | MAE: 3.12 -> 3.24 | Hurt |
| replace CNN encoder with STAR | WB editing · Rendered WB Set 2 | MAE: 3.75 -> 3.67 | Improved |
Closest papers by meaning
- SNR-Aware Low-Light Image Enhancement (CVPR 2022 · similarity 0.86)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · similarity 0.83)
- Star with Bilinear Mapping (CVPR 2025 · similarity 0.83)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.83)
- Specularity Factorization for Low-Light Enhancement (CVPR 2024 · similarity 0.83)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · 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.
