Research map / Image generation / Super-resolution
HIIF: Hierarchical Encoding based Implicit Image Function for Continuous 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.
Proposes HIIF, an implicit image function for continuous arbitrary-scale super-resolution that uses hierarchical positional encoding and multi-head linear attention in the decoder, integrated with EDSR, RDN and SwinIR encoders.
Problem
Existing INR-based continuous super-resolution methods use single-scale positional encoding with MLP decoders, which ignores hierarchical neighborhood structure among local sampling points and limits representation capacity.
Approach
A local hierarchical positional encoding that is progressively injected into successive decoder layers so neighboring sampling points share features at coarser levels, plus a multi-head linear attention block in the implicit decoder to capture non-local information, with a skip connection from bilinear upsampling.
Key results
- HIIF outperforms prior continuous SR methods by up to 0.17 dB PSNR across EDSR, RDN and SwinIR encoders on DIV2K, Set5, Set14, BSD100 and Urban100.
- With EDSR baseline on DIV2K x2, HIIF reaches 34.91 dB vs 34.88 dB for CiaoSR (best prior).
- Ablations on Urban100 with EDSR show removing hierarchical encoding, multi-scale architecture or linear attention each lowers PSNR (e.g., x2: 32.56, 32.47, 32.34 vs 32.69).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HIIF instead of CiaoSR | EDSR baseline continuous super-resolution · Set14 ×4 | PSNR (dB): 28.76 -> 28.81 | Helped |
| Use HIIF instead of CiaoSR | EDSR baseline continuous super-resolution · Urban100 ×4 | PSNR (dB): 26.58 -> 26.57 | Hurt |
| Add hierarchical encoding (ablation) | HIIF with EDSR baseline encoder · Urban100 ×4 | PSNR (dB): 26.38 -> 26.51 | Helped |
| Apply hierarchical encodings progressively at different network layers instead of all at the beginning (ablation) | HIIF with EDSR baseline encoder · Urban100 ×4 | PSNR (dB): 26.35 -> 26.51 | Helped |
| Add multi-head linear attention blocks (ablation) | HIIF with EDSR baseline encoder · Urban100 ×4 | PSNR (dB): 26.25 -> 26.51 | Helped |
Limitations
- Higher inference runtime and memory than MetaSR, LIIF and LTE.
- Gains over prior methods are modest (up to 0.17 dB).
Experimental setup
Datasets: DIV2K
Benchmarks: DIV2K validation Set5 Set14 BSD100 Urban100
Metrics: PSNR
Baselines: MetaSR LIIF LTE CLIT CiaoSR SRNO
Builds on: LIIF LTE EDSR RDN SwinIR
Closest papers by meaning
- Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder (CVPR 2024 · similarity 0.87)
- Auto-Encoded Supervision for Perceptual Image Super-Resolution (CVPR 2025 · similarity 0.86)
- Hierarchical Neural Operator Transformer with Learnable Frequency-aware Loss Prior for Arbitrary-scale Super-resolution (ICML 2024 · similarity 0.86)
- GaussianSR: High Fidelity 2D Gaussian Splatting for Arbitrary-Scale Image Super-Resolution (AAAI 2025 · similarity 0.86)
- Exploring the Low-Pass Filtering Behavior in Image Super-Resolution (ICML 2024 · similarity 0.86)
- Beyond Image Super-Resolution for Image Recognition with Task-Driven Perceptual Loss (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.