Research map / Image generation / GANs
Neural Video Compression Using GANs for Detail Synthesis and Propagation
Fabian Mentzer, Eirikur Agustsson, Johannes Ballé, David Minnen, Nick Johnston, George Toderici
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 GAN loss (ablation) | Ours · MCL-JCV user study | user preference / PSNR: PSNR 35.1 -> 34.5; preferred 68% of the time | Improved |
| add free latent from warped previous reconstruction (ablation) | Ours · MCL-JCV user study | user preference: 23% prefer No free latent over Ours | Improved |
| add conditioning of the discriminator on latents (ablation) | Ours · MCL-JCV user study | user preference: 28% prefer Uncond. Disc. over Ours | Improved |
| add UFlow supervised flow (ablation) | Ours · MCL-JCV user study | user preference: 33% prefer No UFlow over Ours | Improved |
| use bicubic resampling kernel in decoupled scale-space warping (ablation) | MSE-trained DSSW model · MSE R-D | bpp at equal quality: R-D performance improves by 6% | Improved |
| decouple scale-space warping into warp plus adaptive blur | SSW · GPU runtime | speed: 2-3x faster | Improved |
| remove LPIPS distortion loss (use MSE only) | Ours · training | training speed: speeds up training by 35% | Improved |
| raise bitrate of HEVC baseline to 0.14 and 0.22 bpp | HEVC · MCL-JCV user study | user preference: 39% -> 41.2% -> 41.4% prefer HEVC | Mixed |
| condition discriminator on more frames | Ours · MCL-JCV | reconstruction quality | No clear effect |
| replace GAN-trained model by perceptual metrics as predictor of study | LPIPS/PIM/FID · MCL-JCV user study | metric predicts study: predict neural codec studies, not H.264 or HEVC studies | Mixed |
Closest papers by meaning
- High-Fidelity Generative Image Compression (NeurIPS 2020 · similarity 0.90)
- Content-Aware GAN Compression (CVPR 2021 · similarity 0.90)
- Generative Adversarial Networks for Extreme Learned Image Compression (ICCV 2019 · similarity 0.89)
- Multi-level Wavelet-based Generative Adversarial Network for Perceptual Quality Enhancement of Compressed Video (ECCV 2020 · similarity 0.89)
- GAN Compression: Efficient Architectures for Interactive Conditional GANs (CVPR 2020 · similarity 0.89)
- A Style-Based GAN Encoder for High Fidelity Reconstruction of Images and Videos (ECCV 2022 · similarity 0.89)
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.
