AI Paper Atlas

Research map / Image generation / Super-resolution

Towards Bidirectional Arbitrary Image Rescaling: Joint Optimization and Cycle Idempotence

Zhihong Pan, Baopu Li, Dongliang He, Mingde Yao, Wenhao Wu, Tianwei Lin, Xin Li, Errui Ding

CVPR 2022 · Image generation › Super-resolution · approach: Real-world super-resolution · 36 citations (OpenAlex)

Check your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
jointly optimize arbitrary downscaling and upscaling as one bidirectional rescaling modelBAIRNet · arbitrary scales ×1.5 and ×2.5 on Set5, Urban100, and DIV2KPSNR: 40.11 vs 39.78; 36.62 vs 33.99; 37.46 vs 36.60 at ×2.5Improved
use BAIRNet for ×3.5 rescalingBAIRNet† · ×3.5, Set5PSNR: 36.85 vs 37.12Hurt
use BAIRNet for ×3.5 rescalingBAIRNet† · ×3.5, Urban100 and DIV2KPSNR: 32.71 vs 31.84; 33.87 vs 33.84Improved
train with downscaling subpixel weight function (D-SWF) (ablation)BAIRNet · DIV2K, ×4PSNR: 31.19 -> 31.42Improved
weaken LR reference supervision from pixel-level L₂ to Lᶜ₂ or mean-color Lᵐ₂ (ablation)BAIRNet · DIV2K, ×4PSNR: 31.78 -> 32.12; 31.78 -> 32.13Improved
remove LR reference supervision (ablation)BAIRNet · DIV2K, ×4PSNR: 31.78 -> 32.13Improved
use Lᶜ₂ supervision on Cb and Cr channels to suppress false-color artifacts (ablation)BAIRNet · DIV2K, ×4PSNR: 32.13 -> 32.12Hurt
remove scale normalization from the reconstruction loss (ablation)BAIRNet · DIV2K, ×2PSNR: 40.15 -> 39.93Hurt
replace scale-normalized Lₛ reconstruction loss with L₁ (ablation)BAIRNet · DIV2K, ×4PSNR: 32.13 -> 32.14No clear effect
fine-tune using losses over N=3 repeated rescaling cycles (ablation)BAIRNet · DIV2K, ×4, cycles 2–5PSNR improvement over base BAIRNet: 0.16; 0.27; 0.40; 0.54Improved
fine-tune using losses over N=5 repeated rescaling cycles (ablation)BAIRNet · DIV2K, ×4, cycle 1 versus cycles 2–5PSNR improvement over base BAIRNet: 0.05; 0.15; 0.28; 0.44; 0.61Mixed
apply asymmetric-scale fine-tuning (ablation)BAIRNet · asymmetric scales, five benchmark datasetsPSNR: 36.00 -> 36.66 on Set14, ×2/×3Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy