Research map / Vision-language and multimodal / Human and embodied modeling
DreamStruct: Understanding Slides and User Interfaces via Synthetic Data Generation
Yi-Hao Peng, Faria Huq, Yue Jiang, Jason Wu, Xin Yue Li, Jeffrey Bigham, Amy Pavel
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 |
|---|---|---|---|
| train slide element recognizers on synthetic slides instead of equal-sized human-annotated data | slide element recognition · Slide element recognition | average mAP (%): 44.3 -> 55.3 | Improved |
| pretrain UI element recognizers on synthetic UIs before finetuning on 50 human samples | UI element recognition · UI element recognition (50 samples) | average mAP (%): 47.2 -> 52.4 | Improved |
| pretrain on synthetic UIs before finetuning on the full human-annotated set | UI element recognition · UI element recognition (full set) | average mAP (%): 78.8 -> 81.8 | Improved |
| scale synthetic slides for element recognition to twice the human set (ablation) | slide element recognition · Slide element recognition | average mAP (%): 55.3 -> 64.4 | Improved |
| fine-tune VLM on synthetic slides for classification instead of zero-shot | LLaVA / Qwen-VL · Slide classification | average accuracy (%): 46.8 -> 52.2 | Improved |
| fine-tune VLM on small human-annotated data instead of zero-shot | LLaVA / Qwen-VL · Slide classification | average accuracy (%): 46.8 -> 33.8 | Hurt |
| fine-tune VLM on synthetic UIs for classification instead of zero-shot | LLaVA / Qwen-VL · UI classification | average accuracy (%): 38.9 -> 42.3 | Improved |
| fine-tune VLM on small human-annotated UI data instead of zero-shot | LLaVA / Qwen-VL · UI classification | average accuracy (%): 38.9 -> 31.0 | Hurt |
| increase synthetic classification data beyond about 5K samples (ablation) | LLaVA / Qwen-VL · Slide and UI classification | accuracy | No clear effect |
| scale synthetic slides for classification from equal size to 5,033 (ablation) | LLaVA / Qwen-VL · Slide classification | average accuracy (%): 52.2 -> 67.9 | Improved |
| fine-tune LLaVA on synthetic captions instead of existing caption datasets | LLaVA-1.5-13B · Slide captioning | win rate (%): 77.9% win rate | Improved |
| synthetic-trained open model versus GPT-4V for captioning | LLaVA-1.5-13B · Slide captioning | win rate (%): 31.6% win rate | Hurt |
Closest papers by meaning
- AutoPresent: Designing Structured Visuals from Scratch (CVPR 2025 · similarity 0.87)
- JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence (ICLR 2026 · similarity 0.84)
- UI-Vision: A Desktop-centric GUI Benchmark for Visual Perception and Interaction (ICML 2025 · similarity 0.84)
- Rethinking The Training And Evaluation of Rich-Context Layout-to-Image Generation (NeurIPS 2024 · similarity 0.84)
- ShowUI: One Vision-Language-Action Model for GUI Visual Agent (CVPR 2025 · similarity 0.84)
- Widget2Code: From Visual Widgets to UI Code via Multimodal LLMs (CVPR 2026 · similarity 0.84)
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.
