IntReal: 100K Real-World Interleaved Multimodal Samples to Benchmark Unified Models
nanjinguniv · hf · 2026-10-30
Researchers from Nanjing University, Meituan, Imperial College London, and NTU introduce IntReal, a real-world dataset and benchmark for multimodal interleaved generation.
- IntReal-100K: 100k interleaved multimodal samples with over 500k images drawn from 7 real-world sources
- Tasks span Plan (26.1%), Knowledge (21.2%), Tutorial (34.7%), and Story (18.0%) — e.g., travel guides, trail knowledge, cooking visual tutorials
- A data-driven evaluation framework scores text richness, logical consistency, image authenticity, and diversity
The benchmark aims to diagnose unified models' interleaved-generation weaknesses and improve them via targeted training. Project page: yushaohan.github.io/IntReal
More from Multimodal
- Tencent's WeVisCut trains a 9B agent to run the whole editing room, using 33x fewer tokens than Opus — tencent · 2026-10-30
- Hand-drawn-style AI short celebrates Korean Seollal with a good-news magpie — narrator_mochi · 2026-10-30
- Kimi launches 3D Make: connects to Blender and CAD tools to build models, scenes, and games — Kimi_Moonshot · 2026-10-30
- Stunning anime cosplay images made with Qwen Image 2.1 — rarugagamer · 2026-10-30
- Animated short made with ComfyUI: ChatGPT poses, Qwen restyle, MiniMax H3 — bseoan · 2026-10-30
- This Pillow Does Not Exist: a one-click AI pillow design generator — andrew_n_carr · 2026-10-30