Research map / Architectures and efficiency / CNN design
Understanding and Simplifying Perceptual Distances
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 changed | Where | Measured | Effect |
|---|---|---|---|
| Use patch-distribution MMD instead of pixel MSE for super-resolution | Super-resolution model · BSD100, Set5, and Set14, ×4 | SWD: 47.71 -> 26.89; 39.17 -> 24.40; 56.84 -> 18.92 | Helped |
| Use patch-distribution MMD instead of pixel MSE for super-resolution | Super-resolution model · BSD100, Set5, and Set14, ×4 | PSNR: 26.05 -> 23.54; 29.61 -> 26.38; 26.26 -> 23.61 | Hurt |
| Use patch-distribution MMD instead of pixel MSE for super-resolution | Super-resolution model · BSD100, Set5, and Set14, ×4 | NIQE: 5.38 -> 4.55; 6.40 -> 4.67; 5.28 -> 4.38 | Helped |
| Use MMD rather than VGG perceptual loss | Super-resolution model · BSD100, Set5, and Set14, ×4 | SWD: 27.87 -> 26.89; 23.81 -> 24.40; 28.14 -> 18.92 | Mixed |
| Use MMD rather than VGG perceptual loss | Super-resolution model · BSD100, Set5, and Set14, ×4 | NIQE: 7.55 -> 4.55; 7.73 -> 4.67; 7.40 -> 4.38 | Helped |
| Add a local robust patch-comparison term to MMD (MMD++) (ablation) | MMD generalized image mean · GIM with dissimilar target images | visual realism | Helped |
| Use only the deepest VGG layer instead of multiple VGG layers (ablation) | VGG generalized image mean · GIM | visual realism | Hurt |
| Use a random-filter S-CNN rather than low-level metrics | 2AFC image similarity · 2AFC task from [40] | 2AFC score | Helped |
| Average judgments from 20 random networks and constrain first-layer filters to zero mean (ablation) | Random VGG · 2AFC task from [40] | 2AFC score | Helped |
| Replace trained VGG features with random CNN filters (ablation) | GIM and 2AFC image similarity · GIM and 2AFC | perceptual performance | No clear effect |
| Replace deep perceptual feature comparison with patch MMD | Image comparison and prediction losses · GIM and 2AFC | perceptual performance | No clear effect |
| Use MMD alone rather than MMD plus a local robust patch term (ablation) | MMD generalized image mean · GIM with dissimilar target images | visual realism | Hurt |
Closest papers by meaning
- Generating Images with Perceptual Similarity Metrics based on Deep Networks (NeurIPS 2016 · similarity 0.91)
- The Unreasonable Effectiveness of Deep Features as a Perceptual Metric (CVPR 2018 · similarity 0.91)
- The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric (ICLR 2024 · similarity 0.87)
- LipSim: A Provably Robust Perceptual Similarity Metric (ICLR 2024 · similarity 0.85)
- Measuring Per-Unit Interpretability at Scale Without Humans (NeurIPS 2024 · similarity 0.84)
- Discriminating image representations with principal distortions (ICLR 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.