Research map / Video / Video-language understanding
LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models
Yanwei Li, Chengyao Wang, Jiaya Jia
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 instruction-guided context token to a single content token (ablation) | LLaMA-VID · GQA (224px) | accuracy: 53.3 -> 55.5 | Improved |
| add content token to context token (ablation) | LLaMA-VID · GQA (224px) | accuracy: 54.3 -> 55.5 | Improved |
| add one context token to 256 content tokens (ablation) | LLaMA-VID · GQA (224px) | accuracy: 61.9 -> 63.0 | Improved |
| compress content tokens from 256 to 64 (ablation) | LLaMA-VID · GQA (224px) | accuracy: 63.0 -> 60.8 | Hurt |
| compress content tokens from 256 to 1 (ablation) | LLaMA-VID · GQA (224px) | accuracy: 63.0 -> 55.5 | Hurt |
| compress content tokens from 256 to 1 (ablation) | LLaMA-VID · ScienceQA-IMG (224px) | accuracy: 67.7 -> 68.8 | Improved |
| use randomly initialized 2-layer BERT as text decoder instead of none (ablation) | LLaMA-VID · GQA (224px) | accuracy: 53.3 -> 54.1 | Improved |
| use pre-trained QFormer instead of BERT as text decoder (ablation) | LLaMA-VID · GQA (224px) | accuracy: 54.1 -> 55.5 | Improved |
| replace LLaVA-1.5 visual token scheme with context+content tokens | LLaVA-1.5 Vicuna-7B · GQA | accuracy: 62.0 -> 64.3 | Improved |
| scale LLM from Vicuna-7B to 13B | LLaMA-VID · MME | score: 1521.4 -> 1542.3 | Improved |
Later papers that compared with this method
14 later accepted papers report LLaMA-VID as a baseline in their results tables; 14 of them report a better number on most of the results they share with it.
- VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling (ICLR 2026 · better on 1 of 1 shared result)
- B-VLLM: A Vision Large Language Model with Balanced Spatio-Temporal Tokens (ICCV 2025 · better on 7 of 7 shared results)
- Aligning Effective Tokens with Video Anomaly in Large Language Models (ICCV 2025 · better on 4 of 4 shared results)
- Granularity-Adaptive Spatial Evidence Tokenization for Video Question Answering (AAAI 2025 · better on 3 of 3 shared results)
- GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation (CVPR 2025 · better on 2 of 3 shared results)
- Efficient Motion-Aware Video MLLM (CVPR 2025 · better on 1 of 1 shared result)
- HyperGLM: HyperGraph for Video Scene Graph Generation and Anticipation (CVPR 2025 · better on 1 of 1 shared result)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · better on 1 of 1 shared result)
6 more not listed.
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
- VoCo-LLaMA: Towards Vision Compression with Large Language Models (CVPR 2025 · similarity 0.90)
- VISTA-LLAMA: Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens (CVPR 2024 · similarity 0.89)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.88)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.88)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.87)
- Leveraging Textual Memory and Key Frame Reasoning for Full Video Understanding Using Off-the-Shelf LLMs and VLMs (Student Abstract) (AAAI 2025 · 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.
