Research map / Image generation / Autoregressive and latent generation
A Loss Function for Generative Neural Networks Based on Watson’s Perceptual Model
Steffen Czolbe, Oswin Krause, Ingemar Cox, Christian Igel
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 |
|---|---|---|---|
| Use Watson-DFT instead of LPIPS-VGG as the loss | Color-image loss computation · Typical learning scenario | Runtime: $56.8$ s vs. $16.2$ s | Improved |
| Use Watson-DFT instead of LPIPS-VGG as the loss | Color-image loss computation · Typical learning scenario | GPU memory: $2214$ Mb vs. $114$ Mb | Improved |
| Train the VAE with Watson-DFT rather than L2 or SSIM | MNIST-VAE · MNIST | Visual sample quality | Improved |
| Train the VAE with LPIPS-VGG or LPIPS-Squeeze rather than Watson-DFT | MNIST-VAE · MNIST | Visual sample quality | Hurt |
| Train the VAE with Watson-DFT rather than SSIM or Adaptive-Loss | celebA-VAE · celebA | Visual sample quality | Improved |
| Train the VAE with LPIPS-VGG rather than Watson-DFT | celebA-VAE · celebA | Fine-detail generation | Mixed |
| Train the VAE with LPIPS-Squeeze rather than Watson-DFT | celebA-VAE · celebA | Visual artifacts | Hurt |
| Use Watson-DFT rather than L2, SSIM, or Adaptive-Loss | Perceptual similarity metric · 2AFC validation | Agreement with human judgments | Improved |
| Use LPIPS-VGG or LPIPS-Squeeze rather than Watson-DFT | Perceptual similarity metric · 2AFC validation | Agreement with human judgments | Improved |
Closest papers by meaning
- Generating Images with Perceptual Similarity Metrics based on Deep Networks (NeurIPS 2016 · similarity 0.88)
- Variational Lossy Autoencoder (ICLR 2017 · similarity 0.87)
- Perceptual Generative Autoencoders (ICML 2020 · similarity 0.86)
- High-Fidelity Generative Image Compression (NeurIPS 2020 · similarity 0.86)
- Perceptual Losses for Real-Time Style Transfer and Super-Resolution (ECCV 2016 · similarity 0.86)
- JetFormer: An autoregressive generative model of raw images and text (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.
