Research map / Image generation / Super-resolution
GSDD: Generative Space Dataset Distillation for Image Super-resolution
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.
Applies dataset distillation to single image super-resolution by optimizing latent codes of a pre-trained GAN generator (Real-ESRGAN) via GAN inversion and distribution matching, training SR models on ~700 distilled LR-HR pairs.
Problem
Single image super-resolution needs massive LR-HR datasets, which cost training time and storage; coreset selection is not optimized for downstream restoration tasks.
Approach
Distill the OST training set into a small set of synthetic HR/LR pairs by optimizing latent vectors of a pre-trained SISR GAN generator with a distribution-matching loss in a random feature space, plus a regularization term that keeps latent codes paired with real images to avoid homogenization.
Key results
- Models trained on 100 distilled images per class (700 total) reach SSIM/LPIPS comparable to full OST training, with about 93.2% data storage saved and nearly 8x training efficiency for Real-ESRGAN.
- On DPED, GSDD (700 images) scores NIQE 4.7251 vs Real-ESRGAN 4.5896 and beats DGAN and FSSR on NIQE, NRQM, and PI.
- Removing the regularization term (lambda=0) raises LPIPS from 0.1351 to 0.4586 at IPC=100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with 100 distilled images per class instead of the full OST dataset | GSDD with Real-ESRGAN generator · OST / OutdoorSceneTest300 | SSIM: 0.8367 -> 0.7986 | Hurt |
| Train with 100 distilled images per class instead of the full OST dataset | GSDD with Real-ESRGAN generator · OST / OutdoorSceneTest300 | LPIPS: 0.0903 -> 0.1351 | Hurt |
| Use GSDD distilled training data rather than the full training set | Real-ESRGAN · Training efficiency | training efficiency: nearly ×8 | Helped |
| Compress the dataset using GSDD | SISR training data · Data storage | storage space saved: 93.2% | Helped |
| Use 700 GSDD distilled images instead of the comparison methods' training sets | GSDD · DPED | NIQE: 4.5896 -> 4.7251 | Hurt |
| Use 700 GSDD distilled images instead of the comparison methods' training sets | GSDD · DPED | NRQM: 6.1021 -> 5.8498 | Hurt |
| Use 700 GSDD distilled images instead of the comparison methods' training sets | GSDD · DPED | PI: 4.7405 -> 4.4013 | Hurt |
| Add regularization term R during distillation (ablation) | GSDD with Real-ESRGAN generator, IPC 100 · OST / OutdoorSceneTest300 | LPIPS: 0.4586 -> 0.1351 | Helped |
| Set λ = 0.1 instead of λ = 0 (ablation) | GSDD, IPC 100 · OST / OutdoorSceneTest300 | LPIPS: 0.4586 -> 0.1351 | Helped |
| Vary λ across 0∼1 (ablation) | GSDD · OST / OutdoorSceneTest300 | LPIPS | Mixed |
| Set λ = 0.1 rather than other tested values (ablation) | GSDD, IPC 50 and IPC 100 · OST / OutdoorSceneTest300 | LPIPS | Helped |
| Set λ = 0.2 rather than other tested values (ablation) | GSDD, IPC 1 · OST / OutdoorSceneTest300 | LPIPS | Helped |
Surprising findings
- Distilled images look abstract and unidentifiable to humans yet remain useful for training SR networks.
- Distilled data trained with one SR architecture transfers acceptably to other architectures such as SRResNet and SwinIR.
Limitations
- Performance ceiling depends on the quality of the underlying pre-trained SISR GAN generator (Real-ESRGAN results bound GSDD).
- Requires categorical labels in the source dataset for class-wise distribution matching.
- Evaluated mainly on one dataset (OST) with a single scale factor (x4).
Experimental setup
Datasets: OST DPED DIV2K Flickr2K
Benchmarks: OutdoorSceneTest300 DPED
Metrics: SSIM LPIPS NIQE NRQM PI
Baselines: ESRGAN SFTGAN DGAN FSSR RealSR Real-ESRGAN coreset selection SRResNet SwinIR
Builds on: dataset distillation distribution matching GAN inversion Real-ESRGAN differentiable siamese augmentation
Closest papers by meaning
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.88)
- SeD: Semantic-Aware Discriminator for Image Super-Resolution (CVPR 2024 · similarity 0.87)
- Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.87)
- Uncertainty-Aware GAN for Single Image Super Resolution (AAAI 2024 · similarity 0.87)
- Learning Coupled Dictionaries from Unpaired Data for Image Super-Resolution (CVPR 2024 · similarity 0.87)
- Training Generative Image Super-Resolution Models by Wavelet-Domain Losses Enables Better Control of Artifacts (CVPR 2024 · similarity 0.87)
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.