Research map / Video / Video generation
Progressive Growing of Video Tokenizers for Temporally Compact Latent Spaces
Aniruddha Mahapatra, Long Mai, David Bourgin, Yitian Zhang, Feng Liu
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 progressive growing instead of direct full-model training for 8x temporal compression | MagViT-v2 tokenizer · MCL-JCV 8x8x8 z_dim 8 | PSNR / rFVD: 28.47 -> 30.26; 219.36 -> 96.85 | Improved |
| use progressive growing instead of direct training for 16x temporal compression | MagViT-v2 tokenizer · MCL-JCV 8x8x16 z_dim 8 | PSNR / rFVD: 26.15 -> 28.31; 255.51 -> 183.4 | Improved |
| use progressive growing for 8x compression at z_dim 16 | MagViT-v2 tokenizer · MCL-JCV 8x8x8 z_dim 16 | PSNR / rFVD: 28.51 -> 32.06; 135.57 -> 63.64 | Improved |
| use progressive growing for 16x compression at z_dim 16 | MagViT-v2 tokenizer · MCL-JCV 8x8x16 z_dim 16 | PSNR / rFVD: 28.51 -> 30.02; 135.57 -> 115.25 | Improved |
| add residual embedding with AdaNorm and subsampled key-frame encoding on top of naive progressive training (ablation) | ProMAG 8x · MCL-JCV 512x512 | PSNR / rFVD: 30.35 -> 32.06; 88.18 -> 63.64 | Improved |
| naive progressive training (add bottleneck blocks to pretrained 4x model) vs training from scratch (ablation) | MagViT-v2 8x · MCL-JCV 512x512 | PSNR / rFVD: 28.51 -> 30.35; 135.57 -> 88.18 | Improved |
| add GAN loss during progressive growing (8x) (ablation) | ProMAG · MCL-JCV 8x8x8 | PSNR / rFVD: 32.06 -> 31.78; 63.64 -> 53.18 | Mixed |
| add GAN loss during progressive growing (16x, z_dim 16) (ablation) | ProMAG · MCL-JCV 8x8x16 | PSNR / rFVD: 30.02 -> 29.67; 115.25 -> 99.64 | Mixed |
| use 8x model on full frames instead of 4x model on subsampled frames plus 2X interpolation (ablation) | ProMAG · MCL-JCV 512x512 | PSNR / rFVD: 30.68 -> 32.06; 230.16 -> 53.77 | Improved |
| use 16x model instead of 4x model on subsampled frames plus 4X interpolation (ablation) | ProMAG · MCL-JCV 512x512 | PSNR / rFVD: 28.42 -> 29.6; 341.14 -> 147.48 | Improved |
| progressive growing instead of full-model training: training time per iteration (8x) | ProMAG · training efficiency | time per iteration: 0.92s -> 0.42s; 4.25 days -> 1.94 days | Improved |
| progressive growing instead of full-model training: training time (16x) (ablation) | ProMAG · training efficiency | training time: 6.01 days -> 2.21 days | Improved |
Closest papers by meaning
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.89)
- LARP: Tokenizing Videos with a Learned Autoregressive Generative Prior (ICLR 2025 · similarity 0.88)
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.88)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.88)
- Language Model Beats Diffusion - Tokenizer is key to visual generation (ICLR 2024 · similarity 0.88)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · 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.
