Research map / Video / Video-language understanding
PyraTok: Language-Aligned Pyramidal Tokenizer for Video Understanding and Generation
Onkar Susladkar, Tushar Prakash, Adheesh Juvekar, Kiet A. Nguyen, Dong-Hwan Jang, Inderjit S Dhillon, Ismini Lourentzou
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 PyraTok tokenizer instead of prior video VAEs | video reconstruction · WebVid-10M | PSNR: LARP 33.03 -> PyraTok 35.72 | Improved |
| swap native VAE for PyraTok in text-to-video backbones | MotionAura, Open MAGVITv2, OmniGenV2 · WebVid-10M T2V | FVD: 374 -> 365; 433 -> 411; 398 -> 377 | Improved |
| use PyraTok for zero-shot video segmentation instead of OmniTokenizer | discrete video VAE · YouTube-VIS 2021 | mAP: 14.54 -> 24.54 | Improved |
| use PyraTok instead of LARP for zero-shot temporal action localization | discrete video VAE · THUMOS14, ActivityNet v1.3 | avg mAP: 27.42 -> 33.17; 25.53 -> 29.11 | Improved |
| use PyraTok instead of LARP for video classification | discrete video VAE · Kinetics-400 | accuracy (%): 69.27 -> 78.43 | Improved |
| add LaPQ language-aligned pyramidal quantization (ablation) | PyraTok · Reconstruction ablation | PSNR: 31.41 -> 35.72 | Improved |
| add pyramidal multi-scale quantization (ablation) | PyraTok · Reconstruction ablation | PSNR: 34.02 -> 35.72 | Improved |
| add text guidance to quantization (ablation) | PyraTok · Reconstruction ablation | PSNR: 33.43 -> 35.72 | Improved |
| increase quantization blocks from 2 to 4 (ablation) | PyraTok · Reconstruction ablation | PSNR: 33.21 -> 34.78 -> 35.72 | Improved |
| add drift regularization and autoregressive alignment losses (ablation) | PyraTok · Reconstruction ablation | PSNR: 32.17 -> 35.72 | Improved |
| add text-codebook alignment loss (ablation) | PyraTok · Reconstruction ablation | PSNR: 34.11 -> 35.72 | Improved |
| increase codebook size beyond 80K vocabulary (ablation) | PyraTok · Reconstruction | reconstruction quality | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- SweetTok — SweetTok: Semantic-Aware Spatial-Temporal Tokenizer for Compact Video Discretization (ICCV 2025) · better on 1 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
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · similarity 0.89)
- Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024 · similarity 0.88)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.88)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.88)
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.87)
- Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding (CVPR 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.
