Research map / Video / Video segmentation and tracking
ZeroI2V: Zero-Cost Adaptation of Pre-Trained Transformers from Image to Video
Xinhao Li, Yuhan Zhu, 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 |
|---|---|---|---|
| relocate heads (STDHA) instead of temporal channel shift on KV (ablation) | ViT-B/16 temporal modeling · SSv2 val | top-1 accuracy: 64.6 -> 66.0 | Improved |
| apply head relocation to KV only rather than QKV (ablation) | STDHA · SSv2 val | top-1 accuracy: 64.8 -> 66.0 | Improved |
| add Linear Adapter with full adaptation instead of full fine-tuning (ablation) | CLIP ViT-B/16 · SSv2 val | top-1 accuracy: 63.2 -> 66.3 | Improved |
| tune only the attention weights (temporal head) instead of full adaptation (ablation) | Linear Adapter · SSv2 val | top-1 accuracy: 63.2 -> 59.6 vs full fine-tuning | Hurt |
| densely place adapters on all weights (full adaptation) (ablation) | linear and non-linear adapters · SSv2 val | top-1 accuracy: Linear Adapter 64.4 -> 66.0 | Improved |
| use serial Linear Adapter instead of parallel LoRA (ablation) | image-to-video adaptation · SSv2 val | top-1 accuracy: LoRA 65.5 -> Linear Adapter 66.2 | Improved |
| replace linear adapter with non-linear GELU adapter (ablation) | Linear Adapter · SSv2 val | top-1 accuracy: 66.0 vs 66.1 | No clear effect |
| raise adapter bottleneck ratio from 0.25 to 0.5 (ablation) | Linear Adapter · SSv2 val | top-1 accuracy: 66.0 -> 65.8 | No clear effect |
| use more temporal heads (more than 1/6 to 1/4 of heads) (ablation) | STDHA ViT-B · SSv2 val | top-1 accuracy: 66.0 -> 65.6 | Hurt |
| enlarge temporal receptive field beyond 3 for 8 frames (ablation) | STDHA · SSv2 val, 8 frames | top-1 accuracy: 66.0 -> 65.5 | No clear effect |
| enlarge temporal receptive field for 32 frames (ablation) | STDHA · SSv2 val, 32 frames | top-1 accuracy: 67.3 -> 68.6 | Improved |
| use ZeroI2V instead of full fine-tuning of CLIP ViT-L/14 on SSv2 | CLIP ViT-L/14 · SSv2 val | top-1 accuracy: 48.7 -> 70.1 | Improved |
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.
- ST-Adapter — ST-Adapter: Parameter-Efficient Image-to-Video Transfer Learning (NeurIPS 2022) · better on 2 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
- D2ST-Adapter: Disentangled-and-Deformable Spatio-Temporal Adapter for Few-shot Action Recognition (ICCV 2025 · similarity 0.88)
- AIM: Adapting Image Models for Efficient Video Action Recognition (ICLR 2023 · similarity 0.88)
- Revisiting Classifier: Transferring Vision-Language Models for Video Recognition (AAAI 2023 · similarity 0.87)
- Not All Layers Are Equal in Image-to-Video Transfer (NeurIPS 2026 · similarity 0.87)
- Image-to-video Adaptation with Outlier Modeling and Robust Self-learning (AAAI 2025 · similarity 0.86)
- TI2V-Zero: Zero-Shot Image Conditioning for Text-to-Video Diffusion Models (CVPR 2024 · similarity 0.86)
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.
