Research map / Video / Video generation
Keep-or-Drop? Adaptive Tokenizer for Compact Video Representation
Yeonkyeong Lee, Hyunsung Go, Jongmin Kim, Sewoong Lim, Donghoon Lee
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 finer decoder patches than encoder (asymmetric coarse-to-fine decoding) (ablation) | KATok · Panda-70M val reconstruction | PSNR: 29.61 -> 31.26 | Improved |
| raise sparsity loss weight from 0.01 to 0.02 (ablation) | KATok stage 1 · Panda-70M val | PSNR: 28.78 -> 22.46 | Hurt |
| raise sparsity loss weight from 0.01 to 0.02 (ablation) | KATok stage 1 · Panda-70M val | rFVD: 50.38 -> 462.27 | Hurt |
| lower sparsity weight from 0.01 to 0.005 (ablation) | KATok stage 1 · Panda-70M val | PSNR: 28.78 -> 28.67 | No clear effect |
| use KATok adaptive tokenizer instead of OmniTokenizer-VAE under same SiT-XL | SiT-XL diffusion · UCF-101 generation | gFVD: 100.00 -> 61.53 | Improved |
| use KATok instead of ElasticTok-KL under same SiT-XL | SiT-XL diffusion · UCF-101 generation | gFVD: 712.56 -> 61.53 | Improved |
| flip a fraction of ground-truth mask cells before content generation (1-2%) (ablation) | KATok cascaded generation · Kinetics-600 generation | gFVD: 69.58 -> 72.63 (1%), 81.26 (2%) | Hurt |
| flip 10% of mask cells (ablation) | KATok cascaded generation · Kinetics-600 generation | gFVD: 69.58 -> 264.77 | Hurt |
| reduce image patch size from 32^2 to 16^2 (more initial tokens) (ablation) | KATok on ImageNet · ImageNet val reconstruction | PSNR: 25.38 -> 27.94 | Improved |
| use predicted masks from mask prior instead of random masks | cascaded KATok · Kinetics-600 mask FVD | mask FVD: 15,405.39 -> 91.80 | Improved |
Closest papers by meaning
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.88)
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.88)
- CV-VAE: A Compatible Video VAE for Latent Generative Video Models (NeurIPS 2024 · similarity 0.87)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.87)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.87)
- $\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation (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.
