Overview

LMM-10K is a large-scale multimodal video dataset consisting of 10,000 high-quality 4K videos at 60 fps enriched with semantic, perceptual, and content-aware annotations.

• 📹 10,000 sampled 4K video sequences from Pexels
• 🧠 Rich multimodal annotations combining semantic and perceptual information
• 🤖 LLM-generated semantic descriptors using the Qwen3 model
• 🎯 No-reference quality assessment using PIQE, BRISQUE, NIMA, and LIQE
• 📊 Video complexity analysis including SI/TI, VCA , and EVCA
• 🌍 Broad diversity of real-world scenes covering natural, urban, and dynamic content
• ⚙️ Designed to support a wide range of applications:
  • Video enhancement and restoration
  • Content-aware video compression
  • Neural video coding
  • Multimodal learning (vision-language)
  • Generative video modeling
  • Perceptual quality assessment

Citation

@inproceedings{ghasempour2026lmm10k,
  title={{LMM-10K}: {Large}-{Scale} {4K} {Multimodal} {Dataset} for {Perceptual}, {Semantic}, and {Content}-{Aware} {Video} {Processing}},
  author={Ghasempour, Mohammad and Wei, Yiying and Amirpour, Hadi and Timmerer, Christian},
  booktitle={Submitted to ACM Multimedia 2026},
  year={2026}
}