Research map / Image generation / Super-resolution
Binarized Diffusion Model 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 BI-DiffSR, a binarized diffusion model for image super-resolution with a binarization-friendly UNet (CP-Down/CP-Up, CS-Fusion) and timestep-aware redistribution/activation (TaR/TaA).
Problem
Diffusion models for image super-resolution are too costly in memory and computation for edge deployment, and existing binarization methods degrade badly when applied to diffusion UNets because of dimension mismatch, skip-connection fusion difficulty, and activation distribution shifts across timest
Approach
Redesigns the UNet to keep dimensions consistent so identity shortcuts can carry full-precision information (CP-Down/CP-Up, channel-shuffle fusion of skip features), and adds timestep-grouped learnable biases and RPReLU activations (TaR/TaA) that adapt binarized activations to each timestep group at no extra computation.
Key results
- On Manga109 x2, BI-DiffSR reaches 33.99 dB PSNR vs 32.88 dB for the second-best binarization method BBCU.
- Ablation: the full model improves PSNR from 27.66 dB baseline to 32.66 dB on Manga109 x2 (about 5 dB gain).
- BI-DiffSR uses 8.3% of SR3's params and 20.8% of its OPs while achieving 93.6% of SR3's LPIPS on Manga109 x2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the identity shortcut (ablation) | BI-DiffSR baseline · Manga109 ×2 | PSNR (dB): 27.66 -> 29.29 | Helped |
| Add CP-Down and CP-Up (ablation) | BI-DiffSR with identity shortcut · Manga109 ×2 | PSNR (dB): 29.29 -> 31.08 | Helped |
| Add CS-Fusion (ablation) | BI-DiffSR with CP-Down/Up · Manga109 ×2 | PSNR (dB): 31.08 -> 31.99 | Helped |
| Add both TaR and TaA (ablation) | BI-DiffSR with CS-Fusion · Manga109 ×2 | PSNR (dB): 31.99 -> 32.66 | Helped |
| Use CS-Fusion instead of direct addition (ablation) | BI-DiffSR skip-feature fusion · Manga109 ×2 | PSNR (dB): 18.89 -> 31.99 | Helped |
| Use CS-Fusion instead of concatenation (ablation) | BI-DiffSR skip-feature fusion · Manga109 ×2 | PSNR (dB): 31.08 -> 31.99 | Helped |
| Use CS-Fusion instead of split-and-add fusion (ablation) | BI-DiffSR skip-feature fusion · Manga109 ×2 | PSNR (dB): 29.67 -> 31.99 | Helped |
| Use TaR alone instead of no time-aware module (ablation) | BI-DiffSR · Manga109 ×2 | PSNR (dB): 31.99 -> 29.27 | Hurt |
| Use TaA alone instead of no time-aware module (ablation) | BI-DiffSR · Manga109 ×2 | PSNR (dB): 31.99 -> 29.13 | Hurt |
| Use both TaR and TaA rather than TaR alone (ablation) | BI-DiffSR · Manga109 ×2 | PSNR (dB): 29.27 -> 32.66 | Helped |
| Use both TaR and TaA rather than TaA alone (ablation) | BI-DiffSR · Manga109 ×2 | PSNR (dB): 29.13 -> 32.66 | Helped |
| Increase the bias/RPReLU pair count from 1 to 2 (ablation) | BI-DiffSR TaR/TaA · Manga109 ×2 | PSNR (dB): 31.99 -> 32.42 | Helped |
Surprising findings
- Applying only TaR or only TaA degrades PSNR (29.27 and 29.13 dB) relative to no timestep-aware module (31.99 dB); gains appear only when both are used together.
Limitations
- Comparison to full-precision diffusion is limited to SR3 only; no comparison with accelerated samplers or other compressed DMs.
- Still a noticeable gap to full-precision perceptual quality on some metrics (e.g., LPIPS on Urban100 x4 is 0.1823 vs ReActNet's 0.2276 is better, but overall binarized models remain behind SR3 in PSNR).
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 B100 Urban100 Manga109
Metrics: PSNR SSIM LPIPS Params OPs
Baselines: BNN DoReFa XNOR IRNet ReActNet BBCU SR3
Builds on: BNN ReActNet BBCU DDIM SR3
Closest papers by meaning
- BiMaCoSR: Binary One-Step Diffusion Model Leveraging Flexible Matrix Compression for Real Super-Resolution (ICML 2025 · similarity 0.92)
- BiDM: Pushing the Limit of Quantization for Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Flexible Residual Binarization for Image Super-Resolution (ICML 2024 · similarity 0.89)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.89)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.89)
- Efficient Conditional Diffusion Model with Probability Flow Sampling for Image Super-resolution (AAAI 2024 · similarity 0.89)
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.