Research map / Image generation / Super-resolution
Deep Generative Model based Rate-Distortion for Image Downscaling Assessment
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 IDA-RD, a process-based rate-distortion measure that scores image downscaling algorithms by the LPIPS distortion between original HR images and SR reconstructions from deep generative models (SRFlow, PULSE) conditioned on the downscaled output.
Problem
Image downscaling algorithms lack a quantitative evaluation measure because no ground-truth LR images exist and NR-IQA metrics rely on HR-image naturalness that LR images do not have.
Approach
Treat downscaling as the encoder and blind, stochastic super-resolution from a deep generative model as the decoder in a rate-distortion framework; the expected distortion between original HR images and generated HR samples consistent with the downscaled image measures information lost during downscaling, without needing LR ground truth.
Key results
- On synthetic Gaussian blur degradations, IDA-RD increases monotonically with blur strength (0.320 at σ=1.0, 0.434 at σ=2.0, 0.579 at σ=4.0), giving Spearman ρ=1 for the blur series.
- On real-world 4x downscaling, Bilinear (0.129 on DIV2K) scores better than Bicubic (0.157) and nearest neighbour (0.374), and none of the state-of-the-art methods (DPID, L0-regularized, Perceptual) beat Bilinear.
- IDA-RD rises with scaling factor (Bicubic 0.058 at 4x, 0.110 at 8x, 0.228 at 32x on FFHQ), consistent with greater information loss at larger factors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Bilinear instead of Bicubic downscaling | IDA-RD evaluation of classical downscalers · DIV2K, Flickr30K, and RealSR, 4× | IDA-RD: DIV2K: 0.157±0.073 -> 0.129±0.089; Flickr30K: 0.263±0.102 -> 0.239±0.112; RealSR: 0.116±0.052 -> 0.114±0.055 | Helped |
| Increase Gaussian blur strength after bicubic downscaling | synthetic downscaling evaluated by IDA-RD · FFHQ, 8× | IDA-RD: 0.320±0.048 -> 0.434±0.057 -> 0.579±0.065 | Hurt |
| Increase the image scaling factor | Bicubic downscaling evaluated by IDA-RD · FFHQ | IDA-RD: 0.058±0.142 -> 0.110±0.145 -> 0.228±0.056 | Hurt |
| Use SRFlow instead of PULSE as the upscaling function (ablation) | IDA-RD · FFHQ, 4× real-world downscaling | IDA-RD: PULSE: Bicubic 0.171±0.015, Bilinear 0.164±0.015, N.N. 0.254±0.018; SRFlow: Bicubic 0.110±0.145, Bilinear 0.090±0.067, N | Mixed |
| Increase the test-image count NX to 900 (ablation) | IDA-RD with SRFlow · FFHQ synthetic degradations | IDA-RD stability | Helped |
| Increase the number of upscaled samples NQ to 5 (ablation) | IDA-RD with SRFlow · FFHQ, bicubic/bilinear/nearest-neighbor downscaling | IDA-RD stability | Helped |
| Replace LPIPS distortion with PSNR, SSIM, or MS-SSIM (ablation) | IDA-RD · Synthetic degradations | IDA-RD degradation trend | No clear effect |
Surprising findings
- State-of-the-art perceptual downscaling methods (DPID, L0-regularized, Perceptual) do not outperform Bilinear downscaling under IDA-RD, suggesting they lose more information despite better perceptual quality.
- NR-IQA metrics such as NIQE, BRISQUE, MANIQA and CONTRIQUE fail to rank downscaled images correctly, especially at extreme downscaling factors.
Limitations
- SRFlow is unstable on some real-world datasets (DIV2K, Flickr2K), so real-world evaluation is limited to 4x on some datasets and domain-specific datasets elsewhere.
- Results show large standard deviations for some methods and scaling factors, indicating algorithmic bias across individual images.
Experimental setup
Datasets: FFHQ DIV2K Flickr2K Flickr30k RealSR NPRportrait 1.0 AFHQ-Cat
Benchmarks: FFHQ DIV2K Flickr30k RealSR NPRportrait 1.0 AFHQ-Cat
Metrics: IDA-RD score (expected LPIPS) Spearman rank correlation PSNR SSIM MS-SSIM LPIPS
Baselines: NIQE BRISQUE MANIQA CONTRIQUE ESRGAN BSRGAN Real-ESRGAN SR3 RSR
Builds on: SRFlow PULSE LPIPS Shannon rate-distortion theory
Closest papers by meaning
- Block-Based Multi-Scale Image Rescaling (AAAI 2025 · similarity 0.86)
- Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model (CVPR 2025 · similarity 0.85)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.85)
- Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · similarity 0.85)
- Training Generative Image Super-Resolution Models by Wavelet-Domain Losses Enables Better Control of Artifacts (CVPR 2024 · similarity 0.85)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 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.