Research map / Video / Video generation
Nexus-Vid: Efficient Frequency Bridging with Homogeneous Latent Space for Video Unified Models
Jiazheng Xing, Hangjie Yuan, Lingling Cai, Xinyu Liu, Yujie Wei, Fei Du, Tao Feng, Hai Ci, Jiasheng Tang, Weihua Chen, Fan Wang, Yong Liu
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 |
|---|---|---|---|
| Hand off to a large Wan2.1-14B generator at inference via UPFB | Omni-Video · VBench-T2V | Total score: 83.82 -> 84.12 | Improved |
| Same change, semantic score | Omni-Video · VBench-T2V | Semantic score: 79.10 -> 80.52 | Improved |
| Same change on reasoning benchmark | Omni-Video · VR-Bench | Total score: 72.78 -> 79.28 | Improved |
| Swap large generator to Wan2.2-A14B | Lumos-Nexus · VR-Bench | Total score: 80.98 -> 81.90 | Improved |
| Use temporal gating sharpness gamma_w 0.3 rather than 0.5 (ablation) | UPFB · VR-Bench | Total score: 75.10 -> 79.28 | Improved |
| Vary gamma_w on VBench total (ablation) | UPFB · VBench-T2V | Total score: 84.02 to 84.12 | No clear effect |
| Use moderate bandwidth schedule (0.35, 0.70) instead of wide (1.00, 2.00) (ablation) | UPFB · VBench-T2V | Total score: 83.56 -> 84.12 | Improved |
| Use moderate bandwidth schedule instead of narrow (0.05, 0.10) (ablation) | UPFB · VBench-T2V | Total score: 83.99 -> 84.12 | Improved |
| Add RMS alignment and energy re-balancing (ablation) | UPFB · VBench-T2V | Total score: 84.07 -> 84.12 | Improved |
| Use frequency-aware progressive bridging instead of Direct Add averaging (ablation) | Lumos-Nexus · VBench-T2V | Total score: 83.28 -> 84.12 | Improved |
| Same, on VR-Bench (ablation) | Lumos-Nexus · VR-Bench | Total score: 77.33 -> 79.28 | Improved |
| Run both generators at inference | Wan2.1-14B · Inference cost | Latency: 35.1 -> 40.5 s/step | Hurt |
Closest papers by meaning
- From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models (ECCV 2026 · similarity 0.89)
- TiViBench: Benchmarking Think-in-Video Reasoning for Video Generation (CVPR 2026 · similarity 0.89)
- CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models (NeurIPS 2026 · similarity 0.88)
- OpenCoF: Learning to Reason Through Video Generation (NeurIPS 2026 · similarity 0.86)
- Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment (NeurIPS 2026 · similarity 0.86)
- SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation (ICLR 2025 · similarity 0.85)
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.
