Research map / Image generation / Super-resolution
PocketSR: The Super-Resolution Expert in Your Pocket Mobiles
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.
Builds PocketSR, a 146M-parameter single-step diffusion model for real-world super-resolution that replaces the SD VAE with a lightweight LiteED, prunes the U-Net via online annealing with multi-layer feature distillation, and runs 4K images in 0.8 s.
Problem
Large generative diffusion models for real-world image super-resolution are too computationally expensive and slow for edge and mobile deployment.
Approach
Replace the heavy SD VAE with an ultra-compact encoder-decoder (LiteED, with dual-path feature injection and adaptive skip connections), and prune the SD-Turbo U-Net by gradually annealing original modules into lightweight parallel replacements, supervised by multi-layer feature distillation from the unpruned teacher.
Key results
- PocketSR has 146M parameters and 225 GMACs, with 0.016 s inference at 512x512 (62.5 FPS) on a single GPU, versus 0.11 s for OSEDiff and 0.03 s for AdcSR.
- On RealSR, PocketSR achieves the best LPIPS (0.2713) and NIQE (5.067) among compared methods, with PSNR 25.47 matching AdcSR.
- On DRealSR, PocketSR reports the best LPIPS (0.2962), DISTS (0.2139) and NIQE (5.809), with NIQE about 10% better than the second-best method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Adaptive Skip Connections (ASC) to LiteED encoder (ablation) | LiteED · RealSR | LPIPS: 0.2737 -> 0.2474 | Helped |
| Add Dual-path Feature Injection (DFI) to LiteED (ablation) | LiteED · RealSR | LPIPS: 0.2580 -> 0.2474 | Helped |
| Replace PixelUnshuffle with the Lite encoder (ablation) | LiteED · RealSR | LPIPS: 0.2697 -> 0.2474 | Helped |
| Replace Lite decoder with the SD decoder (ablation) | LiteED · RealSR | PSNR: 25.61 -> 25.95 | Helped |
| Use online annealing pruning instead of direct lightweight module replacement (ablation) | PocketSR U-Net · RealSR | NIQE | Helped |
| Add multi-layer feature distillation during pruning (ablation) | PocketSR U-Net · RealSR | LPIPS: +3.8% | Helped |
| Use multi-layer rather than single-layer output distillation (ablation) | PocketSR U-Net · RealSR | LPIPS | Helped |
| Initialize the pruned U-Net with the Stable Diffusion prior (ablation) | PocketSR · RealSR | PSNR: 24.66 -> 25.47 | Helped |
| Prune residual blocks at depths III and IV instead of no residual-block pruning (ablation) | PocketSR U-Net · RealSR | LPIPS: 0.2653 -> 0.2695 | Hurt |
| Prune cross-attention at all depths instead of no cross-attention pruning (ablation) | PocketSR U-Net · RealSR | LPIPS: 0.2653 -> 0.2666 | Mixed |
| Prune self-attention at depth IV (ablation) | PocketSR U-Net · RealSR | LPIPS: 0.2653 -> 0.2677 | Hurt |
| Prune FFNs at depths III and IV instead of no FFN pruning (ablation) | PocketSR U-Net · RealSR | LPIPS: 0.2653 -> 0.2695 | Hurt |
Surprising findings
- Pruning deeper U-Net blocks (depth III and IV) has negligible impact on perceptual quality, while pruning shallower blocks causes large degradation.
- Removing the SD prior and training the pruned U-Net from scratch lowers PSNR by about 0.8 dB on RealSR, though it gives a slightly better NIQE score.
Limitations
- Detail generation under severe degradations remains limited.
- The framework has not been optimized for or tested on edge hardware.
Experimental setup
Datasets: LSDIR FFHQ RealSR DRealSR DIV2K validation
Benchmarks: RealSR DRealSR DIV2K validation
Metrics: LPIPS DISTS PSNR SSIM NIQE MUSIQ FID MACs parameters inference time
Baselines: StableSR DiffBIR SeeSR ResShift SinSR OSEDiff AdcSR
Builds on: SD-Turbo Stable Diffusion VAE OSEDiff AdcSR TAESD
Closest papers by meaning
- Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning (CVPR 2025 · similarity 0.89)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.86)
- Thinking in Granularity: Dynamic Quantization for Image Super-Resolution by Intriguing Multi-Granularity Clues (AAAI 2025 · similarity 0.85)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.85)
- Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.85)
- SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training (CVPR 2025 · 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.