Research map / Image generation / Super-resolution
Low-Light Face Super-resolution via Illumination, Structure, and Texture Associated Representation
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 IC-FSRDENet, a two-step framework for joint low-light low-resolution face super-resolution: IC-FSRNet (illumination-correction SR with iterative illumination estimation and SR) followed by a diffusion-based detail enhancement network (DENet).
Problem
Recovering high-resolution face images from low-light, low-resolution inputs, where cascaded FSR and low-light enhancement pipelines accumulate errors and fail to recover credible facial structure and texture.
Approach
Decomposes the task into structural fidelity maintenance (IC-FSRNet, which jointly performs bilateral-grid illumination estimation and SR with mutual iterative refinement via cross-attention) and texture consistency learning (a conditional DDPM-based DENet that takes the coarse SR output as side information).
Key results
- IC-FSRNet reaches PSNR 21.68 and SSIM 0.6349 on CelebAMask-HQ, about 3.04 dB above the best cascaded solution, with 8.5M parameters.
- IC-FSRDENet improves LPIPS to 0.1267 and NIQE to 6.129 versus IC-FSRNet's 0.1990 and 10.585, at PSNR 20.53.
- Ablation: iterative illumination-SR joint learning raises PSNR from 20.39 (Model 1) to 21.68 dB; DENet trained directly on LLR inputs gets the worst PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the joint IC-FSRNet instead of the strongest cascaded baseline | IC-FSRNet · CelebAMaskHQ test | PSNR: 18.64 -> 21.68 | Helped |
| use the joint IC-FSRNet instead of the strongest cascaded baseline | IC-FSRNet · CelebAMaskHQ test | SSIM: 0.5749 -> 0.6349 | Helped |
| use the joint IC-FSRNet instead of the strongest cascaded baseline | IC-FSRNet · CelebAMaskHQ test | LPIPS: 0.2022 -> 0.1990 | Helped |
| add diffusion-based detail enhancement after IC-FSRNet (ablation) | IC-FSRNet · CelebAMaskHQ test | NIQE: 10.585 -> 6.129 | Helped |
| add diffusion-based detail enhancement after IC-FSRNet (ablation) | IC-FSRNet · CelebAMaskHQ test | LPIPS: 0.1990 -> 0.1267 | Helped |
| add diffusion-based detail enhancement after IC-FSRNet (ablation) | IC-FSRNet · CelebAMaskHQ test | PSNR: 21.68 -> 20.53 | Hurt |
| add illumination estimation to the RCAB FSR baseline (ablation) | Model 1 · CelebAMaskHQ test | PSNR: 20.39 -> 21.48 | Helped |
| add illumination estimation to the RCAB FSR baseline (ablation) | Model 1 · CelebAMaskHQ test | LPIPS: 0.2466 -> 0.2018 | Helped |
| use iterative rather than one-step illumination estimation (ablation) | Model 2 · CelebAMaskHQ test | PSNR: 21.48 -> 21.68 | Helped |
| use iterative rather than one-step illumination estimation (ablation) | Model 2 · CelebAMaskHQ test | SSIM: 0.6310 -> 0.6349 | Helped |
| replace diffusion enhancement with GAN-based enhancement (ablation) | IC-FSRNet · CelebAMaskHQ test | PSNR: 21.68 -> 20.46 | Hurt |
| replace diffusion enhancement with GAN-based enhancement (ablation) | IC-FSRNet · CelebAMaskHQ test | LPIPS: 0.1990 -> 0.1417 | Hurt |
Surprising findings
- Training DENet directly on LLR-HR pairs without IC-FSRNet yields the best NIQE but the worst PSNR, with color deviation and artifacts.
Limitations
- Diffusion-based DENet makes inference lengthy.
- Complex coupled degradations limit recoverable facial attributes; extreme cases can yield incorrect attributes.
- Evaluated only on simulated degradations of a single dataset (CelebAMask-HQ).
Experimental setup
Datasets: CelebAMask-HQ
Benchmarks: CelebAMask-HQ
Metrics: PSNR SSIM NIQE LPIPS cosine similarity (face recognition via DeepFace)
Baselines: SISN SCTANet SFMNet FECNet LLformer LEDNet FSR+LLIE cascades LLIE+FSR cascades IC-FSRGAN
Builds on: DDPM Zero-DCE SR3 (Saharia et al. 2021) RCAB
Closest papers by meaning
- PortraitSR: Artist-Inspired Prior Learning for Progressive Face Super-Resolution (AAAI 2026 · similarity 0.88)
- Super-Resolving Very Low-Resolution Face Images With Supplementary Attributes (CVPR 2018 · similarity 0.88)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.86)
- SFDM: Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images (CVPR 2025 · similarity 0.86)
- Blind Face Restoration under Extreme Conditions: Leveraging 3D-2D Prior Fusion for Superior Structural and Texture Recovery (AAAI 2024 · similarity 0.85)
- Unveiling Details in the Dark: Simultaneous Brightening and Zooming for Low-Light Image Enhancement (AAAI 2024 · similarity 0.85)
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.