Research map / Image generation / Image restoration
Exposure Bracketing Is All You Need For A High-Quality Image
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 BracketIRE/BracketIRE+, a recurrent network (TMRNet) with frame-specific aggregation modules plus self-supervised real-image adaptation, to jointly denoise, deblur, HDR-reconstruct and super-resolve from 5 exposure-bracketed raw images, trained on synthetic HDR-video pairs and 200 smartphone night scenes.
Problem
Obtaining noise-free, blur-free, high dynamic range, high-resolution photos in low light; existing multi-image methods address only subsets of these degradations and generalize poorly to real captures without ground truth.
Approach
Jointly process multi-exposure bracketed raw images for denoising, deblurring, HDR and SR using a temporally modulated recurrent network whose aggregation module has shared common and frame-specific parts, and adapt to real unlabeled images with a self-supervised loss using temporal pseudo-targets plus an EMA regularizer.
Key results
- TMRNet without adaptation reaches 39.35 dB PSNR on the BracketIRE synthetic set, 0.25 dB above RBSR (39.10 dB).
- Self-supervised adaptation raises real-world CLIPIQA/MANIQA from 0.2003/0.2181 to 0.2537/0.2422 on BracketIRE.
- Step-by-step processing (37.93 dB) is inferior to joint processing (39.35 dB) on the synthetic BracketIRE task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add self-supervised real-image adaptation to TMRNet (ablation) | TMRNet · BracketIRE real-world | CLIPIQA: 0.2003 -> 0.2537 | Helped |
| Add self-supervised real-image adaptation to TMRNet (ablation) | TMRNet · BurstSR real-world test | PSNR: 44.70 -> 45.57 | Helped |
| Remove the EMA regularization loss while retaining the self-supervised loss (ablation) | TMRNet real-image adaptation · BracketIRE and BracketIRE+ real-world | training stability | Hurt |
| Use EMA regularization alone without the self-supervised loss (ablation) | TMRNet real-image adaptation · BracketIRE real-world | CLIPIQA: 0.2003 -> 0.2003 | No clear effect |
| Increase the self-supervised loss weight from λ_self=1 to λ_self=2 (ablation) | TMRNet real-image adaptation · BracketIRE+ real-world | CLIPIQA: 0.3676 -> 0.3815 | Helped |
| Increase the self-supervised loss weight from λ_self=1 to λ_self=4 (ablation) | TMRNet real-image adaptation · BracketIRE real-world | CLIPIQA: 0.2537 -> 0.1974 | Hurt |
| Add frame-specific aggregation blocks to a common-only network (ablation) | TMRNet · BracketIRE synthetic | PSNR: 38.66 -> 39.35 | Helped |
| Add common aggregation blocks to a specific-only network (ablation) | TMRNet · BracketIRE synthetic | PSNR: 38.91 -> 39.35 | Helped |
| Increase specific aggregation depth from 24 to 32 blocks while keeping common depth at 16 (ablation) | TMRNet · BracketIRE synthetic | PSNR: 39.35 -> 39.36 | No clear effect |
| Increase common aggregation depth from 16 to 24 blocks while keeping specific depth at 24 (ablation) | TMRNet · BracketIRE synthetic | PSNR: 39.35 -> 39.31 | No clear effect |
| Use joint denoising, deblurring, and HDR reconstruction instead of three sequential models (ablation) | TMRNet · BracketIRE synthetic | PSNR: 37.93 -> 39.35 | Helped |
| Use five exposure-bracketed images rather than five same-exposure Y2 burst images (ablation) | TMRNet · BracketIRE synthetic | PSNR: 35.05 -> 39.35 | Helped |
Surprising findings
- Removing the EMA regularizer makes self-supervised adaptation collapse to NaN, and the EMA loss alone leaves parameters unchanged.
- Higher lambda_self (2 or 4) sometimes gives better no-reference scores but reduces contrast and visual quality, so the no-reference metrics are treated as only auxiliary.
Limitations
- Requires sensor-specific training because noise model parameters differ across sensors; generalization to other cameras is limited.
- TMRNet has higher FLOPs (20,040 G) than the Transformer-based baselines.
- Synthetic blur from frame interpolation (RIFE) still leaves a gap with real blur.
Experimental setup
Datasets: HDM HDR video dataset (Froehlich et al.) Kalantari HDR dataset BurstSR
Benchmarks: BracketIRE synthetic test set (290 pairs) BracketIRE+ synthetic test set BracketIRE real-world test set (100 scenes) Kalantari HDR dataset BurstSR
Metrics: PSNR SSIM LPIPS CLIPIQA MANIQA
Baselines: DBSR MFIR BIPNet Burstormer RBSR AHDRNet HDRGAN HDR-Transformer SCTNet Kim et al.
Builds on: RBSR EDVR UPI Noise2Noise
Closest papers by meaning
- UltraLED: Learning to See Everything in Ultra-High Dynamic Range Scenes (NeurIPS 2025 · similarity 0.84)
- Self-Supervised High Dynamic Range Imaging with Multi-Exposure Images in Dynamic Scenes (ICLR 2024 · similarity 0.83)
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.83)
- Hybrid-Supervised Dual-Search: Leveraging Automatic Learning for Loss-Free Multi-Exposure Image Fusion (AAAI 2024 · similarity 0.83)
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.83)
- VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook (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.