Research map / Video / Video-language understanding
MoVA: Learning Asymmetric Dual Projections for Modular Long Video-Text Alignment
Peiyuan Zhu, Shaoan Xie, Zijian Li, Yifan Shen, Namrata Deka, Harsh Shrivastava, Guangyi Chen, Kun Zhang
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 temporal mask network and concept mask network (dual asymmetric projections) vs global loss only (ablation) | CLIP-based video-text model · ActivityNet T2V | R@1: 43.7 -> 47.6 | Improved |
| use dual projections instead of SmartCLIP-style single-direction visual-to-text masking (ablation) | MoVA · ActivityNet T2V | R@1: 35.6 -> 47.6 | Improved |
| apply temporal masks on the global text representation rather than raw tokens (ablation) | MoVA · ActivityNet T2V | R@1: 42.0 -> 47.6 | Improved |
| add DSL post-processing | MoVA · ActivityNet T2V | R@1: 47.8 -> 53.6 | Improved |
| remove different-frame same-mask loss (ablation) | MoVA · ActivityNet T2V | R@1: -3.6 R@1 when removed | Hurt |
| use soft masks instead of hard masks (ablation) | MoVA · ActivityNet T2V | R@1: 47.8 -> 47.4 | No clear effect |
| increase frames from 128 to 160 (ablation) | MoVA · ActivityNet T2V | R@1: 47.8 -> 46.6 | Hurt |
| increase TMN depth beyond a shallow depth (ablation) | MoVA · ActivityNet T2V | R@1: 47.8 (2 blocks) -> 47.0 (8 blocks) | No clear effect |
| set global-to-modular loss ratio above 1.2 (ablation) | MoVA · loss weighting sweep | retrieval | Hurt |
| use MoVA text encoder in VideoCrafter2 instead of CLIP | VideoCrafter2 · VBench | average score: 0.8456 -> 0.8708 | Improved |
| MoVA vs VideoCLIP-XL on short-caption DiDeMo | MoVA · DiDeMo T2V | R@1: 38.6 -> 57.5 | Improved |
Closest papers by meaning
- VideoComp: Advancing Fine-Grained Compositional and Temporal Alignment in Video-Text Models (CVPR 2025 · similarity 0.90)
- Text as Partial Constraint: Core–Residual Alignment for Robust Vision–Language Learning (NeurIPS 2026 · similarity 0.90)
- TOPA: Extending Large Language Models for Video Understanding via Text-Only Pre-Alignment (NeurIPS 2024 · similarity 0.89)
- SmartCLIP: Modular Vision-language Alignment with Identification Guarantees (CVPR 2025 · similarity 0.88)
- DiscoVLA: Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval (CVPR 2025 · similarity 0.88)
- CovMatch: Cross-Covariance Guided Multimodal Dataset Distillation with Trainable Text Encoder (NeurIPS 2025 · similarity 0.88)
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.
