TLDR: This research introduces the first public dataset for slide animation modeling, consisting of 12,000 text-JSON-video triplets covering all PowerPoint effects. It also presents a LoRA-fine-tuned Qwen-2.5-VL-7B model that significantly outperforms existing VLMs and closed-source models (like GPT-4.1 and Gemini-2.5-Pro) in understanding slide animations. Additionally, a new evaluation metric, CODA (Coverage–Order–Detail Assessment), is proposed to rigorously assess action coverage, temporal order, and detail fidelity. This work provides a robust benchmark and foundation for future VLM-based dynamic slide generation.
In today’s fast-paced world, presentations are a cornerstone of communication, whether in education, business, or science. Slide animations, like fade-ins or fly-ins, are crucial for keeping audiences engaged and delivering information effectively. However, most AI tools designed for creating slides still lack the ability to handle these dynamic animations. This is largely because there hasn’t been a public dataset available for training AI models on slide animations, and existing visual-language models (VLMs) struggle with understanding the timing and sequence of these effects.
A new research paper, titled “Animation Needs Attention: A Holistic Approach to Slides Animation Comprehension with Visual-Language Models”, addresses this significant gap. The researchers, Yifan Jiang, Yibo Xue, Yukun Kang, Pin Zheng, Jian Peng, Feiran Wu, and Changliang Xu, have introduced a groundbreaking solution to enable AI to better understand and potentially generate slide animations.
The core of their work is the release of the first-ever public dataset specifically for slide animation modeling. This extensive dataset comprises 12,000 unique sets, each containing a natural-language description of an animation, a technical JSON file detailing the animation, and a rendered video of the animation in action. This comprehensive collection covers every built-in animation effect available in PowerPoint, making it an invaluable resource for AI training.
To leverage this new dataset, the team fine-tuned a powerful visual-language model called Qwen-2.5-VL-7B. They used a technique called Low-Rank Adaptation (LoRA), which allows for efficient training by adding only a small number of new trainable components while keeping most of the original model frozen. This method proved highly effective, significantly boosting the model’s ability to grasp fine-grained motion cues and maintain the correct temporal order of animations.
Also Read:
- Benchmarking Vision-Language Models: Unpacking Performance Across Diverse AI Tasks
- AI’s New Frontier: Detecting Road Crashes with Language Models
A New Way to Measure Success
Recognizing that traditional evaluation metrics don’t fully capture the nuances of animation understanding, the researchers also proposed a new metric called Coverage–Order–Detail Assessment (CODA). This innovative, AI-based metric evaluates three key aspects of an animation description: action coverage (how much of the animation is described), temporal order (whether the sequence of events is correct), and detail fidelity (how accurately the specific parameters of each animation are captured). CODA provides a more comprehensive way to assess how well an AI model understands slide animations.
The results of their experiments were impressive. The LoRA-enhanced Qwen-2.5-VL-7B model consistently outperformed leading models like GPT-4.1 and Gemini-2.5-Pro on various metrics, including BLEU-4, ROUGE-L, SPICE, and all CODA sub-scores. On a manually created test set of slides, the LoRA model showed remarkable improvements, demonstrating its ability to generalize beyond the synthetic data it was trained on.
This research marks a significant step forward in making AI-driven slide generation tools more dynamic and engaging. By providing both a much-needed dataset and an improved model, this work lays a strong foundation for future advancements in AI’s ability to understand and create animated presentations. For more in-depth information, you can read the full research paper. The paper also discusses limitations, such as the semantic richness of static slides and computational resource constraints, pointing towards exciting avenues for future research, including more sophisticated page composition logic and advanced temporal modeling in visual encoders.
