Research map / Video / Video generation
Spotlight: Identifying and Localizing Video Generation Errors Using VLMs
Aditya Aravind Chinchure, Sahithya Ravi, Pushkar Shukla, Vered Shwartz, Leonid Sigal
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 |
|---|---|---|---|
| use multi-agent per-error-type prompting (ablation) | Qwen3-VL-8B · Spotlight (tau=0.7) | S+P score M: 0.148 -> 0.254 | Improved |
| use sliding-window 2s sub-clips at 4 fps (ablation) | Qwen3-VL-8B · Spotlight (tau=0.7) | S+P score M: 0.148 -> 0.189 | Improved |
| split into reason-then-localize stages (ablation) | Qwen3-VL-8B · Spotlight (tau=0.7) | S+P score M: 0.148 -> 0.189 | Improved |
| use multi-agent prompting instead of sequential decomposition (ablation) | Qwen3-VL-8B · Spotlight (tau=0.7) | S+P score M: 0.189 -> 0.254 | Improved |
| use multi-agent prompting (ablation) | Qwen3-VL-8B · Spotlight (tau=0.7) | Recall M%: 73.1% -> 61.7% | Hurt |
| use multi-agent prompting (ablation) | Qwen3-VL-8B · Spotlight (tau=0.7) | Precision M%: 27.1% -> 49.8% | Improved |
| scale Qwen3-VL from 8B to 30B-A3B | Qwen3-VL zero-shot · Spotlight (tau=0.7) | S+P score M: 0.148 -> 0.137 | No clear effect |
| use Gemini 2.5 Pro instead of Qwen3-VL zero-shot | zero-shot VLM · Spotlight (tau=0.7) | S+P score M: 0.148 -> 0.250 | Improved |
| human annotators instead of best VLM | zero-shot Gemini 2.5 Pro · Spotlight (tau=0.7) | S+P score M: 0.250 -> 0.508 | Improved |
| use multi-agent prompting on appearance/disappearance errors | Qwen3-VL-8B · Spotlight S+P, App/Disapp | S+P score M: 0.099 -> 0.017 | Hurt |
| use multi-agent prompting on physics errors vs Gemini | Qwen3-VL-8B · Spotlight S+P, Physics | S+P score M: 0.149 -> 0.205 | Improved |
Closest papers by meaning
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.90)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.89)
- ViLMA: A Zero-Shot Benchmark for Linguistic and Temporal Grounding in Video-Language Models (ICLR 2024 · similarity 0.88)
- CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models (NeurIPS 2026 · similarity 0.88)
- ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (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.
