Research map / Video / Video generation
Hierarchical Patch VAE-GAN: Generating Diverse Videos from a Single Sample
Shir Gur, Sagie Benaim, Lior Wolf
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 hp-VAE-GAN trained on one video instead of multi-sample video GANs (MoCoGAN) | hp-VAE-GAN · SVFID vs 1st nearest neighbor | mean SVFID: 19.0 -> 6.1 | Improved |
| use hp-VAE-GAN vs TGAN when compared to a 2nd nearest neighbor video | hp-VAE-GAN · SVFID vs 2nd NN, TGAN samples | mean SVFID: 20.6 -> 21.0 | Hurt |
| use hp-VAE-GAN instead of multi-sample baselines | hp-VAE-GAN · UCF-101 test | FID: 42.0 -> 22.1 | Improved |
| use patch-VAE for coarse scales plus patch-GAN instead of SinGAN/ConSinGAN 3D extension | hp-VAE-GAN · single-video user study | diversity MOS: 2.9 -> 3.8 | Improved |
| use hp-VAE-GAN instead of SinGAN on single-image generation | hp-VAE-GAN · single-image user study | realness MOS: 2.3 -> 3.5 | Improved |
| increase number of patch-VAE levels M (ablation) | hp-VAE-GAN · single video generation | diversity vs SVFID | Mixed |
| use only VAE levels (M=N=9) (ablation) | hp-VAE-GAN · single video generation | visual quality | Hurt |
| use a single VAE level (M=1) (ablation) | hp-VAE-GAN · single video generation | diversity | Hurt |
| train all levels instead of freezing earlier networks (ablation) | hp-VAE-GAN · single video generation | diversity and SVFID: diversity -51%, SVFID +680% | Hurt |
| set decoder receptive field r=1 (ablation) | patch-VAE · M=5 patch-VAE output | sample quality (qualitative) | Hurt |
| set encoder receptive field r=1 (ablation) | patch-VAE · M=5 patch-VAE output | diversity (qualitative) | Hurt |
| replace our patch-VAE by PatchVAE of Gupta et al. (ablation) | patch-VAE · M=5 patch-VAE output | sample quality (qualitative) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MoCoGAN — MoCoGAN: Decomposing Motion and Content for Video Generation (CVPR 2018) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Diverse Generation from a Single Video Made Possible (ECCV 2022 · similarity 0.90)
- Generating Diverse High-Fidelity Images with VQ-VAE-2 (NeurIPS 2019 · similarity 0.89)
- Diverse Rare Sample Generation with Pretrained GANs (AAAI 2025 · similarity 0.87)
- SF-V: Single Forward Video Generation Model (NeurIPS 2024 · similarity 0.86)
- CV-VAE: A Compatible Video VAE for Latent Generative Video Models (NeurIPS 2024 · similarity 0.86)
- Hierarchical Patch Diffusion Models for High-Resolution Video Generation (CVPR 2024 · similarity 0.86)
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.
