Research map / Image generation / Image restoration
Deep Generative Models for Distribution-Preserving Lossy Compression
Michael Tschannen, Eirikur Agustsson, Mario Lucic
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 |
|---|---|---|---|
| Combine the WAE objective with a WGAN term to train the generator (ablation) | Wasserstein++ versus WAE · CelebA | rFID: 38.55 -> 10.93 | Improved |
| Combine the WAE objective with a WGAN term to train the generator (ablation) | Wasserstein++ versus WAE · CelebA | MSE: 0.0165 -> 0.0277 | Hurt |
| Combine the WAE objective with a WGAN term to train the generator (ablation) | Wasserstein++ versus WAE · LSUN bedrooms | rFID: 42.59 -> 27.52 | Improved |
| Combine the WAE objective with a WGAN term to train the generator (ablation) | Wasserstein++ versus WAE · LSUN bedrooms | MSE: 0.0099 -> 0.0321 | Hurt |
| Use Wasserstein++ rather than WGAN-GP to train the generator (ablation) | DPLC · CelebA | sFID: 22.70 -> 23.36 | Hurt |
| Use Wasserstein++ rather than WGAN-GP to train the generator (ablation) | DPLC · LSUN bedrooms | sFID: 45.52 -> 60.97 | Hurt |
| Use Wasserstein++ rather than WAE to train the generator (ablation) | DPLC · CelebA | MSE | Hurt |
| Use WGAN-GP rather than Wasserstein++ to train the generator (ablation) | DPLC · LSUN bedrooms | reconstruction FID | Hurt |
| Train with the regularized distortion-plus-divergence objective instead of the two-stage DPLC procedure (ablation) | Generative compression · CelebA and LSUN bedrooms | conditional pixel variance | Hurt |
| Use a GAN-based regularizer instead of MMD to match the prior (ablation) | DPLC latent mapping · Low bitrates | mode coverage | Hurt |
| Increase the depth of B to address blurry WGAN-GP reconstructions (ablation) | DPLC with WGAN-GP generator · LSUN bedrooms | reconstruction quality | No clear effect |
| Increase lambda_MMD to reduce mismatch between B(E(X)) and the prior (ablation) | DPLC with WGAN-GP generator · LSUN bedrooms | reconstruction quality | No clear effect |
Closest papers by meaning
- High-Fidelity Generative Image Compression (NeurIPS 2020 · similarity 0.91)
- Generative Adversarial Networks for Extreme Learned Image Compression (ICCV 2019 · similarity 0.90)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.88)
- Lossy Image Compression with Conditional Diffusion Models (NeurIPS 2023 · similarity 0.88)
- LOSSY COMPRESSION WITH DISTRIBUTION SHIFT AS ENTROPY CONSTRAINED OPTIMAL TRANSPORT (ICLR 2022 · similarity 0.87)
- Improving Statistical Fidelity for Neural Image Compression with Implicit Local Likelihood Models (ICML 2023 · similarity 0.87)
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.
