Research map / Image generation / Fast diffusion sampling
Efficient Diffusion as Low Light Enhancer
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.
Distills a multi-step diffusion low-light enhancer into 2/4/8-step students by refining the teacher trajectory with a reflectance-aware residual space and linear score extrapolation (ReDDiT).
Problem
Diffusion-based low-light image enhancement needs many iterative sampling steps; existing acceleration methods (DDIM, progressive and consistency distillation) degrade performance significantly when steps are reduced.
Approach
Identifies fitting error and inference gap as causes of degradation; mitigates fitting error by linearly interpolating the teacher score functions (scaling parameter omega) and the inference gap by shifting the diffusion space to a residual space built from a Retinex-based reflectance estimate (RATR module), then distills the refined teacher trajectory into a student with trajectory-matching plus
Key results
- Ours-8 reaches 28.090 PSNR on LOLv1 and Ours-4 reaches 31.250 PSNR on LOLv2-real, new SOTA over GSAD and WCDM.
- Ours-2 achieves 27.397 PSNR on LOLv1 and 30.613 on LOLv2-real, comparable to 10-step prior diffusion methods.
- Reports 2-step inference at 13.1 FPS versus 2.62 FPS for GSAD (10 steps) with 17.43M parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ReDDiT-8 instead of GSAD | Diffusion LLIE · LOLv1 | PSNR: 27.839 -> 28.090 | Helped |
| Use ReDDiT-4 instead of WCDM | Diffusion LLIE · LOLv2-real | PSNR: 30.461 -> 31.250 | Helped |
| Use ReDDiT-8 instead of Retinexformer | Diffusion LLIE · LOLv2-synthetic | PSNR: 29.040 -> 30.166 | Helped |
| Add reflectance-aware trajectory refinement (ablation) | ReDDiT · LOLv1, LOLv2-real, and LOLv2-synthetic | PSNR | Helped |
| Add reflectance-aware trajectory refinement (ablation) | ReDDiT · LOLv2-real | PSNR | Helped |
| Use reflectance rather than clean images as the trajectory-refinement component (ablation) | ReDDiT · LOLv1, LOLv2-real, and LOLv2-synthetic | PSNR | Helped |
| Use reflectance rather than low-light images as the trajectory-refinement component (ablation) | ReDDiT · LOLv1, LOLv2-real, and LOLv2-synthetic | PSNR | Helped |
| Use PairLIE output rather than reflectance as the trajectory-refinement component (ablation) | ReDDiT · LOLv1, LOLv2-real, and LOLv2-synthetic | PSNR | Hurt |
| Remove the perceptual loss (ablation) | ReDDiT · LOLv1, LOLv2-real, and LOLv2-synthetic | PSNR | Hurt |
| Remove the pixel loss (ablation) | ReDDiT · LOLv1, LOLv2-real, and LOLv2-synthetic | PSNR | Hurt |
| Vary refinement strength ω (ablation) | ReDDiT · LOLv1, LOLv2-real, and LOLv2-synthetic | PSNR | Mixed |
| Reduce sampling steps from 8 to 4 (ablation) | ReDDiT · LOLv2-real | PSNR: 30.919 -> 31.250 | Helped |
Surprising findings
- Distilled student can surpass its multi-step teacher: 4-step model exceeds the 8-step diffusion model on LOLv2-real after refinement.
Limitations
- Single-step restoration yields low PSNR and artifacts, so one-step is not yet viable.
- Lightweight denoising network was not explored.
- Teacher-student setup depends on a non-learned denoiser and Retinex illumination estimate for the reflectance-aware refinement.
Experimental setup
Datasets: LOLv1 LOLv2 SID SDSD
Benchmarks: LOLv1 LOLv2-real LOLv2-synthetic SID SDSD DICM LIME MEF NPE VV
Metrics: PSNR SSIM LPIPS NIQE
Baselines: PyDiff WCDM GSAD DDIM PD CD LLFlow Retinexformer SNR-Net LLFormer
Builds on: progressive distillation consistency distillation DDIM GSAD Retinex theory
Closest papers by meaning
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.89)
- Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement (AAAI 2024 · similarity 0.88)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.87)
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.87)
- Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network (NeurIPS 2026 · similarity 0.86)
- One More Step: A Versatile Plug-and-Play Module for Rectifying Diffusion Schedule Flaws and Enhancing Low-Frequency Controls (CVPR 2024 · similarity 0.86)
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.