Research map / Image generation / Autoregressive and latent generation
Shift-tolerant Perceptual Similarity Metric
Abhijay Ghildyal, Feng Liu
Read the paperCheck 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 changed | Where | Measured | Effect |
|---|---|---|---|
| add BlurPool anti-aliasing and reduce conv-1 stride to 2 (ablation) | LPIPS (Alex) · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 2.87 -> 1.46 | Improved |
| reduce conv-1 stride from 2 to 1 on top of BlurPool (ablation) | LPIPS (Alex) · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 1.46 -> 0.66 | Improved |
| add F-Conv (full convolution) (ablation) | LPIPS (Alex) · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 2.87 -> 2.78 | No clear effect |
| use learned adaptive BlurPool instead of fixed BlurPool (ablation) | LPIPS (Alex) · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 1.46 -> 2.80 | Hurt |
| add BlurPool to pooling layers (reflection pad 1) (ablation) | VGG-16 LPIPS · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 3.01 -> 0.66 | Improved |
| add BlurPool to pooling layers (reflection pad 1) (ablation) | SqueezeNet LPIPS · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 7.41 -> 2.03 | Improved |
| add BlurPool to pooling layers (reflection pad 1) (ablation) | SqueezeNet LPIPS · BAPPS 256x256 | 2AFC: 69.61 -> 69.24 | Hurt |
| use larger reflection padding (2 instead of 1) in BlurPool (ablation) | ResNet-18 LPIPS · BAPPS 256x256 | 2AFC: 69.95 -> 70.14 | Improved |
| place BlurPool before activation when conv-1 stride is 1 (ablation) | LPIPS (Alex) · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 0.66 -> 0.57 | Improved |
| place BlurPool before activation when conv-1 stride is 2 (ablation) | LPIPS (Alex) · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 1.46 -> 2.06 | Hurt |
| add shift-tolerant design (BlurPool, stride 1, blur before activation) | LPIPS (Alex) · BAPPS 256x256 shifted | rank-flip rate (1 pixel): 2.81 -> 0.57 | Improved |
| add shift-tolerant design | LPIPS (Alex) · CLIC | accuracy (%): 73.68 -> 76.97 | Improved |
Closest papers by meaning
- LipSim: A Provably Robust Perceptual Similarity Metric (ICLR 2024 · similarity 0.88)
- The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric (NeurIPS 2026 · similarity 0.88)
- When does perceptual alignment benefit vision representations? (NeurIPS 2024 · similarity 0.87)
- The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric (ICLR 2024 · similarity 0.86)
- Decoupling Semantic Similarity from Spatial Alignment for Neural Networks. (NeurIPS 2024 · similarity 0.85)
- ReSi: A Comprehensive Benchmark for Representational Similarity Measures (ICLR 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.
