Research map / Image generation / Image restoration
Plug-and-Play Tri-Branch Invertible Block for Image Rescaling
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 a tri-branch invertible block (T-InvBlock) that splits the low-frequency branch into YCbCr luminance and chrominance branches, plus an all-zero high-frequency mapping, and plugs it into IRN and SAIN to form T-IRN and T-SAIN for image rescaling.
Problem
Image rescaling (joint downscaling and upscaling) with invertible neural networks suffers from channel redundancy when low-frequency components are processed in RGB, and relies on distribution modeling (Gaussian or GMM) for high-frequency components, which limits HR reconstruction quality and robust
Approach
Decompose the low-frequency Haar branch into separate luminance (Y) and chrominance (CbCr) branches within an invertible block, and replace the sampled high-frequency latent with an all-zero tensor during upscaling so that reconstruction relies on the LR image content itself.
Key results
- T-IRN at x2 improves Y-channel PSNR by roughly 0.9 to 1.3 dB over IRN across Set5, Set14, BSD100, Urban100 and DIV2K (e.g., DIV2K 45.46 vs 44.32 dB).
- T-SAIN at x2 under JPEG QF 30-90 on DIV2K gains 0.4 to 0.6 dB PSNR over SAIN.
- At x4 under JPEG QF 30-90, T-SAIN improves PSNR by 0.2 to 0.5 dB over SAIN while using fewer parameters (5.45M vs 6.49M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace IRN’s dual-branch block and Gaussian sampling with T-InvBlock and all-zero mapping | IRN (T-IRN) · Set5, Set14, BSD100, Urban100, and DIV2K val, ×2 | PSNR: 43.99 -> 44.86; 40.79 -> 41.70; 41.32 -> 42.68; 39.92 -> 41.05; 44.32 -> 45.46 | Helped |
| Replace IRN’s dual-branch block and Gaussian sampling with T-InvBlock and all-zero mapping | IRN (T-IRN) · Set5, Set14, BSD100, Urban100, and DIV2K val, ×4 | PSNR: 36.19 -> 36.29; 32.67 -> 32.70; 31.64 -> 31.64; 31.41 -> 31.19; 35.07 -> 35.10 | Mixed |
| Replace SAIN’s block and GMM sampling with T-InvBlock and all-zero mapping | SAIN (T-SAIN) · DIV2K val, JPEG QF=30, 50, 70, 80, 90, ×2 | PSNR: 31.47 -> 31.89; 33.17 -> 33.71; 34.73 -> 35.20; 35.46 -> 35.86; 35.96 -> 36.30 | Helped |
| Replace SAIN’s block and GMM sampling with T-InvBlock and all-zero mapping | SAIN (T-SAIN) · DIV2K val, JPEG QF=30, 50, 70, 80, 90, ×4 | PSNR: 27.90 -> 28.08; 29.05 -> 29.43; 29.83 -> 30.34; 30.13 -> 30.69; 30.31 -> 30.92 | Helped |
| Convert low-frequency RGB to YCbCr while retaining the V-InvBlock (ablation) | IRN · Set14, Urban100, DIV2K, ×2 | PSNR: 40.79 -> 40.47; 39.92 -> 39.33; 44.32 -> 44.02 | Hurt |
| Use an all-zero high-frequency tensor instead of Gaussian sampling (ablation) | IRN · Set14, Urban100, DIV2K, ×2 | PSNR: 40.79 -> 41.37; 39.92 -> 40.55; 44.32 -> 45.18 | Helped |
| Convert the low-frequency branch to YCbCr (ablation) | SAIN · DIV2K val, JPEG QF=50, 70, 80, ×4 | PSNR: 29.05 -> 29.28; 29.83 -> 30.13; 30.13 -> 30.47 | Helped |
| Replace GMM high-frequency modeling with an all-zero tensor (ablation) | SAIN · DIV2K val, JPEG QF=50, 70, 80, ×4 | PSNR: 29.05 -> 29.15; 29.83 -> 29.87; 30.13 -> 30.12 | Mixed |
Surprising findings
- Replacing the high-frequency GMM/Gaussian sampling with an all-zero tensor performs comparably or better than sampling, as IRN ablations and SAIN's learned near-zero-mean GMM suggest.
- Simply converting RGB low-frequency to YCbCr without the T-InvBlock did not significantly improve IRN (Set14 40.47 vs 40.79 dB, Urban100 and DIV2K decreased).
Limitations
- At x4 scale, T-IRN does not surpass IRN on all datasets (e.g., Set14 SSIM, Urban100 PSNR lower).
- At QF=80 the all-zero strategy gives a slight decline for T-SAIN ablation.
- Compute hardware is not stated in the provided text.
Experimental setup
Datasets: DIV2K
Benchmarks: Set5 Set14 BSD100 Urban100 DIV2K validation
Metrics: PSNR SSIM
Baselines: IRN SAIN SRCNN EDSR RDN RCAN SAN ESRGAN CAR TAD/TAU CNN-CR/CNN-SR HCFlow
Builds on: IRN SAIN Haar wavelet transform affine coupling layer
Closest papers by meaning
- Block-Based Multi-Scale Image Rescaling (AAAI 2025 · similarity 0.89)
- AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · similarity 0.85)
- Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range Tone Mapping (CVPR 2024 · similarity 0.85)
- Deep Generative Model based Rate-Distortion for Image Downscaling Assessment (CVPR 2024 · similarity 0.85)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.84)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.84)
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.