Research map / Image generation / Autoregressive and latent generation
Generating Images with Perceptual Similarity Metrics based on Deep Networks
Read the paperCheck 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 |
|---|---|---|---|
| replace image-space SE/l1 loss with DeePSiM feature-plus-adversarial loss for autoencoder training | autoencoder on STL-10 · STL-10 reconstruction | normalized Euclidean reconstruction error: 15.3 -> 19.8/21.5 | Hurt |
| use DeePSiM loss (Exemplar-CNN or AlexNet features) instead of SE/l1 for autoencoder features | autoencoder hidden representation used for linear SVM · STL-10 classification | classification accuracy: 34.6 -> 50.1 (Exemplar-CNN), 52.3 (AlexNet) | Helped |
| use AlexNet conv3 comparator instead of Exemplar-CNN conv3 for autoencoder loss (ablation) | DeePSiM autoencoder · STL-10 reconstruction | normalized Euclidean reconstruction error: 19.8 -> 21.5 | Hurt |
| ablate image-space loss, adversarial loss, and feature loss in feature inversion of FC6 (ablation) | DeePSiM AlexNet inversion · ILSVRC-2012 validation FC6 inversion | visual quality | No clear effect |
| add adversarial loss to feature-space loss (ablation) | feature inversion of AlexNet · ILSVRC-2012 validation FC6 inversion | qualitative image realism | Helped |
| use random-weight AlexNet or VideoNet conv5 instead of AlexNet conv5 as comparator (ablation) | DeePSiM inversion of FC6 · ILSVRC-2012 validation FC6 inversion | qualitative reconstruction quality | Hurt |
| use discriminator features as comparator instead of AlexNet conv5 (ablation) | DeePSiM inversion · ILSVRC-2012 validation | qualitative reconstruction quality | Hurt |
| add adversarial loss (natural image prior) to VAE trained with feature loss instead of SE | variational autoencoder on ILSVRC-2012 · VAE samples | qualitative realism of samples | Helped |
| use feature-space loss with the comparator instead of image-space loss for autoencoder/inversion | DeePSiM · Iterative re-encoding of AlexNet features | feature preservation over iterations | Helped |
| Adding a generator input noise to produce multiple pre-images per feature (ablation) | DeePSiM feature inversion · sampling pre-images | diversity of reconstructions | No clear effect |
Closest papers by meaning
- Understanding and Simplifying Perceptual Distances (CVPR 2021 · similarity 0.91)
- The Unreasonable Effectiveness of Deep Features as a Perceptual Metric (CVPR 2018 · similarity 0.90)
- The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric (ICLR 2024 · similarity 0.85)
- Towards flexible perception with visual memory (ICML 2025 · similarity 0.84)
- LipSim: A Provably Robust Perceptual Similarity Metric (ICLR 2024 · similarity 0.84)
- The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric (NeurIPS 2026 · 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.