Research map / Architectures and efficiency / Quantization
BiMaCoSR: Binary One-Step Diffusion Model Leveraging Flexible Matrix Compression for Real 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.
Combines 1-bit binarization with one-step distillation for diffusion-based real-world super-resolution, adding a low-rank SVD-initialized branch and a sparse outlier branch to recover full-precision information.
Problem
Diffusion-based super-resolution is too heavy in memory and computation for edge devices; naive binarization of diffusion models causes catastrophic collapse, and multi-step sampling remains slow.
Approach
Binarize the one-step SinSR backbone and add two full-precision auxiliary branches to each binarized layer: a low-rank branch (LoRA-like, initialized from top-r SVD components of the weight) carrying low-frequency information, and a sparse branch that keeps the k largest-magnitude outlier weights in full precision with fixed coordinates; the binarized branch then models high-frequency detail.
Key results
- On RealSR, BiMaCoSR (1-bit, one step) reports PSNR 26.84, LPIPS 0.3375 and FID 86.09, beating all listed binarization baselines (best baseline XNOR/ReActNet FID about 105.7/109.4).
- Compared with the FP one-step SinSR, the paper reports a 23.8x compression ratio and 27.4x speedup (FLOPs 1.83G vs 50.23G, params 4.98M vs 118.59M).
- Rank-8 LRMB and sparse skip SMB initialization give the best trade-off in ablations; SVD init of LRMB improves SSIM, LPIPS and CLIP-IQA+ over LoRA-style init.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LRMB to BMB (ablation) | BiMaCoSR · RealSR | PSNR: 26.41 -> 26.95 | Helped |
| add SMB to BMB + LRMB (ablation) | BiMaCoSR · RealSR | LPIPS: 0.3400 -> 0.3375 | Helped |
| add SMB to BMB + LRMB (ablation) | BiMaCoSR · RealSR | PSNR: 26.95 -> 26.84 | Hurt |
| use distillation loss instead of SinSR loss (ablation) | BiMaCoSR · RealSR | PSNR: 26.37 -> 26.83 | Helped |
| initialize LRMB with SVD instead of Zero + Random (ablation) | BiMaCoSR · RealSR | LPIPS: 0.3497 -> 0.3375 | Helped |
| set LRMB rank to 4 rather than 8 (ablation) | BiMaCoSR · RealSR | PSNR: 26.84 -> 26.29 | Hurt |
| set LRMB rank to 12 rather than 8 (ablation) | BiMaCoSR · RealSR | PSNR: 26.84 -> 26.97 | Helped |
| set LRMB rank to 16 rather than 8 (ablation) | BiMaCoSR · RealSR | CLIP-IQA+: 0.4800 -> 0.4971 | Helped |
| initialize SMB with sparse skip instead of Zero initial (ablation) | BiMaCoSR · RealSR | PSNR: 26.72 -> 26.84 | Helped |
| initialize SMB with sparse skip instead of Uni-shortcut (ablation) | BiMaCoSR · RealSR | PSNR: 25.69 -> 26.84 | Helped |
| binarize the one-step SinSR model with BiMaCoSR branches | SinSR · RealSR | PSNR: 26.51 -> 26.84 | Helped |
| binarize the one-step SinSR model with BiMaCoSR branches | SinSR · DRealSR | PSNR: 27.89 -> 29.33 | Helped |
Surprising findings
- The LRMB rank-16 variant performs best only on CLIP-IQA+, with larger rank hurting other metrics and making training unstable.
- BiMaCoSR can surpass the full-precision SinSR and ResShift on PSNR and LPIPS on RealSR.
Limitations
- Performance on PSNR and SSIM drops slightly with the SMB branch relative to LRMB-only.
- Trained with only 100K iterations at 64x64 crops on one A6000 and evaluated on a small set of benchmarks.
- Gap between BiMaCoSR and the FP one-step model remains visible, and it is only evaluated for 4x SR.
Experimental setup
Datasets: ImageNet RealESRGAN degradation pipeline
Benchmarks: RealSR DRealSR DIV2K-Val
Metrics: PSNR SSIM LPIPS MUSIQ MANIQA DISTS FID NIQE CLIP-IQA+ FLOPs
Baselines: XNOR ReActNet BBCU ReSTE BiDM SinSR ResShift
Builds on: SinSR ResShift XNOR-Net LoRA
Closest papers by meaning
- Binarized Diffusion Model for Image Super-Resolution (NeurIPS 2024 · similarity 0.92)
- Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning (CVPR 2025 · similarity 0.89)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.89)
- BiDM: Pushing the Limit of Quantization for Diffusion Models (NeurIPS 2024 · similarity 0.88)
- BinaryDM: Accurate Weight Binarization for Efficient Diffusion Models (ICLR 2025 · similarity 0.87)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 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.