Research map / Video / Video generation
InternVideo-Next: Towards World-Understanding Video Models
Chenting Wang, Yuhan Zhu, Yicheng Xu, Jiange Yang, Ziang Yan, Yali Wang, Yi Wang, Limin Wang
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 |
|---|---|---|---|
| naively add SigLIP alignment to pixel reconstruction with a linear decoder (ablation) | Stage 1 · K400 linear probing | top-1 accuracy: 70.7 -> 69.8 | Hurt |
| add semantic alignment to pixel reconstruction baseline (ablation) | Stage 1 · K400 linear probing | top-1 accuracy: 47.2 -> 69.8 | Improved |
| replace linear decoder with conditional diffusion decoder (ablation) | Stage 1 Pixel Rec. + Align · K400 linear probing | top-1 accuracy: 69.8 -> 74.2 | Improved |
| add text-decoder predictor initialization without diffusion decoder (ablation) | Stage 1 Pixel Rec. + Align · K400 linear probing | top-1 accuracy: 69.8 -> 69.4 | Hurt |
| combine diffusion decoder with text-decoder init (ablation) | Stage 1 · K400 linear probing | top-1 accuracy: 74.2 -> 75.8 | Improved |
| use diffusion MLP decoder (D6 W1536) instead of linear head (ablation) | Stage 1 · K400 linear probing | top-1 accuracy: 69.4 -> 75.8 | Improved |
| initialize predictor from pretrained text decoder (ablation) | Stage 1 predictor (ModernBert-L last5) · K400 linear probing | top-1 accuracy: 74.2 -> 75.8 | Improved |
| use SigLIP image teacher instead of DinoV2 or CLIP-ViT for semantic alignment (ablation) | Stage 1 · K400 linear probing | top-1 accuracy: 69.3 -> 75.8 (DinoV2 vs SigLIP) | Improved |
| add Stage 2 latent prediction with frozen Stage 1 target (ablation) | InternVideo-Next ViT-B · SSv2 linear probing | top-1 accuracy: 36.9 -> 56.9 | Improved |
| use momentum-updated target instead of frozen target in Stage 2 (ablation) | Stage 2 · SSv2 linear probing | top-1 accuracy: 56.9 -> 54.3 | Hurt |
| use frozen SigLIP2 target instead of Stage 1 target in Stage 2 (ablation) | Stage 2 · SSv2 linear probing | top-1 accuracy: 56.9 -> 45.7 | Hurt |
| add pixel reconstruction loss in Stage 2 (ablation) | Stage 2 · SSv2 linear probing | top-1 accuracy: 56.9 -> 57.0 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VideoMAEv2 — VideoMAE V2: Scaling Video Masked Autoencoders With Dual Masking (CVPR 2023) · better on 1 of 1 shared result
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
- Video Language Model Pretraining with Spatio-temporal Masking (CVPR 2025 · similarity 0.88)
- A Simple Recipe for Contrastively Pre-training Video-First Encoders Beyond 16 Frames (CVPR 2024 · similarity 0.88)
- Masked Diffusion Vision-Language Models for Temporal Action Localization (NeurIPS 2026 · similarity 0.87)
- FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video Translation (CVPR 2024 · similarity 0.87)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.87)
- Bernini: Latent Semantic Planning for Video Diffusion (NeurIPS 2026 · 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.
