Research map / Image generation / Super-resolution
Flexible Residual Binarization 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.
Proposes Flexible Residual Binarization (FRB), combining second-order residual weight binarization (SRB) and distillation-guided binarization training (DBT), to binarize CNN (FRBC) and Transformer (FRBT) image super-resolution networks.
Problem
Binarized image super-resolution suffers large performance drops because binarizing weights and activations loses high-frequency information and distorts pixel-wise representation content.
Approach
Binarize the weight reconstruction residual in a second order (SRB) so binarized weights better approximate full-precision ones, and align block-level normalized representations of the binary student with a full-precision teacher (DBT) via an L2-normalized distillation loss added to the L1 pixel loss.
Key results
- FRBC (×2, 1-bit W/A) reaches 37.71 dB on Set5 and 31.15 dB on Urban100, outperforming ReActNet (37.26 / 30.85) and BBCU-L (37.58 / 31.12).
- FRBT at ×4 reaches 31.79 dB on Set5 and 25.55 dB on Urban100 vs SwinIR-S full precision 32.44 / 26.47, with about 89.6% parameter and 93.1% ops reduction.
- Ablation (×2, SRResNet backbone): SRB alone improves DoReFa on Urban100 from 29.15 to 30.56 dB; full FRB reaches 30.74 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add second-order residual binarization (ablation) | DoReFa-binarized SRResNet · B100 ×2 | PSNR: 31.25 -> 31.77 | Helped |
| Add second-order residual binarization (ablation) | DoReFa-binarized SRResNet · Urban100 ×2 | PSNR: 29.15 -> 30.56 | Helped |
| Add second-order residual binarization (ablation) | DoReFa-binarized SRResNet · Manga109 ×2 | PSNR: 35.66 -> 37.51 | Helped |
| Add distillation-guided binarization training (ablation) | DoReFa-binarized SRResNet · B100 ×2 | PSNR: 31.25 -> 31.26 | No clear effect |
| Add distillation-guided binarization training (ablation) | DoReFa-binarized SRResNet · Urban100 ×2 | PSNR: 29.15 -> 29.18 | No clear effect |
| Add distillation-guided binarization training (ablation) | DoReFa-binarized SRResNet · Manga109 ×2 | PSNR: 35.66 -> 35.77 | Helped |
| Combine SRB and DBT (ablation) | DoReFa-binarized SRResNet · B100 ×2 | PSNR: 31.25 -> 31.83 | Helped |
| Combine SRB and DBT (ablation) | DoReFa-binarized SRResNet · Urban100 ×2 | PSNR: 29.15 -> 30.74 | Helped |
| Combine SRB and DBT (ablation) | DoReFa-binarized SRResNet · Manga109 ×2 | PSNR: 35.66 -> 37.64 | Helped |
| Binarize SRResNet with FRBC instead of BBCU-L | SRResNet · Urban100 ×4 | PSNR: 25.62 -> 25.61 | No clear effect |
| Binarize SwinIR S with FRBT | SwinIR S · Urban100 ×2 | PSNR: 32.76 -> 31.13 | Hurt |
Surprising findings
- Transformer binarization (FRBT) shows a larger performance gap to full precision than CNN binarization (FRBC), and FRBT uses far fewer parameters and ops than FRBC.
- DBT alone gives only marginal gains over DoReFa (except about 0.11 dB on Manga109), suggesting generic knowledge distillation provides little benefit for binarized SR without SR-specific design.
Limitations
- Binarized Transformer (FRBT) has a notable performance gap to its full-precision SwinIR-S counterpart.
- Compression and speedup are theoretical estimates; real training and inference acceleration is not demonstrated.
- Only the body of the network is binarized; head and tail stay full precision, so overall compression is limited.
Experimental setup
Datasets: DIV2K
Benchmarks: Set5 Set14 B100 Urban100 Manga109
Metrics: PSNR SSIM Params Ops
Baselines: BNN DoReFa Bi-Real IRNet BAM BTM ReActNet BBCU-L
Builds on: XNOR-Net DoReFa BAM BTM ReActNet
Closest papers by meaning
- Binarized Diffusion Model for Image Super-Resolution (NeurIPS 2024 · similarity 0.89)
- 2DQuant: Low-bit Post-Training Quantization for Image Super-Resolution (NeurIPS 2024 · similarity 0.87)
- DegBins: Degradation-Driven Binning for Depth Super-Resolution (NeurIPS 2026 · similarity 0.86)
- AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · similarity 0.85)
- Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning (CVPR 2025 · similarity 0.85)
- BiMaCoSR: Binary One-Step Diffusion Model Leveraging Flexible Matrix Compression for Real Super-Resolution (ICML 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.