Research map / Video / Video generation
DeRA: Decoupled Representation Alignment for Video Tokenization
Pengbo Guo, Junke Wang, Zhen Xing, Chengxu Liu, Daoguo Dong, Xueming Qian, Zuxuan Wu
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 DINOv3 alignment to appearance stream (ablation) | DeRA base tokenizer · UCF-101 | rFVD: 24.47 -> 20.14 | Improved |
| add InternVideo2 alignment to motion stream (ablation) | DeRA base tokenizer · UCF-101 | rFVD: 24.47 -> 22.64 | Improved |
| align both streams with separate foundation models (ablation) | DeRA base tokenizer · UCF-101 | rFVD: 24.47 -> 18.83 | Improved |
| align both streams (ablation) | DeRA base tokenizer · UCF-101 | gFVD: 112 -> 94 | Improved |
| use DINOv3 instead of DINOv2 as image target (ablation) | DeRA · UCF-101 | rFVD: 23.98 -> 20.14 | Improved |
| use InternVideo2 instead of VideoMAEv2 as video target (ablation) | DeRA · UCF-101 | rFVD: 24.14 -> 22.64 | Improved |
| add SACP gradient reformulation (ablation) | DeRA · UCF-101 | rFVD: 19.81 -> 18.83 | Improved |
| add soft gradient-conflict loss with weight 0.5 (ablation) | DeRA · UCF-101 | gFVD: 97 -> 96 | Improved |
| add soft gradient-conflict loss with weight 1 (ablation) | DeRA · UCF-101 | gFVD: 97 -> 103 | Hurt |
| two-stage alignment instead of joint one-stage (ablation) | DeRA · UCF-101 | rFVD: 18.83 -> 19.52 | No clear effect |
| raise lambda_m from 0.5 to 1 (ablation) | DeRA · UCF-101 | rFVD: 18.83 -> 18.94 | No clear effect |
| align deeper encoder layer (depth 6 to 12) with DINOv3 (ablation) | DeRA · UCF-101 | rFVD: 21.34 -> 20.14 | Improved |
Closest papers by meaning
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.87)
- LARP: Tokenizing Videos with a Learned Autoregressive Generative Prior (ICLR 2025 · similarity 0.87)
- OmniTokenizer: A Joint Image-Video Tokenizer for Visual Generation (NeurIPS 2024 · similarity 0.85)
- Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024 · similarity 0.85)
- End-to-End Vision Tokenizer Tuning (NeurIPS 2025 · similarity 0.85)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.84)
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.
