Research map / Image generation / Autoregressive and latent generation
VideoMAETok: Boosting Video Diffusion Models via Masked Autoencoders as Tokenizers
Zhan Tong, Tinne Tuytelaars
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 high-ratio token masking (ablation) | VideoMAETok S-B · UCF-101, SiT-B | gFVD: 418.8 -> 220.8 | Improved |
| add interpolative latent noise (ablation) | VideoMAETok S-B · UCF-101, SiT-B | gFVD: 418.8 -> 178.8 | Improved |
| combine masking and latent noise (ablation) | VideoMAETok S-B · UCF-101, SiT-B | gFVD: 418.8 -> 106.2 | Improved |
| combine masking and latent noise (ablation) | VideoMAETok S-B · Kinetics-400 val reconstruction | rFVD: 6.0 -> 19.8 | Hurt |
| increase reconstruction loss weight 10x (ablation) | VideoMAETok S-B · UCF-101, SiT-B | gFVD: 106.2 -> 3998.2 | Hurt |
| increase latent channels from 16 to 32 (ablation) | VideoMAETok S-B · UCF-101, SiT-B | gFVD: 106.2 -> 3948 | Hurt |
| increase latent channels from 4 to 16 (ablation) | VideoMAETok S-B · UCF-101, SiT-B | gFVD: 1427 -> 106.2 | Improved |
| use causal attention instead of non-causal (ablation) | VideoMAETok S-B · UCF-101, SiT-B | gFVD: 106.2 -> 125.4 | Hurt |
| scale tokenizer from ViT S-S to B-L (ablation) | VideoMAETok · UCF-101, SiT-B | gFVD: 172.9 -> 87.7 | Improved |
| scale generator from SiT-B to SiT-XL (ablation) | VideoMAETok S-B · UCF-101 | gFVD: 106.2 -> 82.2 | Improved |
| use heavier decoder than encoder (asymmetric design) (ablation) | VideoMAETok · UCF-101, SiT-B | gFVD: 172.9 (S-S) -> 106.2 (S-B) | Improved |
| raise masking ratio from 0.8 to 0.9 (ablation) | VideoMAETok S-B · UCF-101, SiT-B | gFVD: 106.2 -> degraded at 0.9 | Hurt |
Closest papers by meaning
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.90)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.90)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.90)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.90)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.89)
- $\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation (NeurIPS 2025 · 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.
