The explosive growth of visual content, especially short-form videos, together with rapid advances in computer vision has fueled the large-scale deployment of vision models across edge devices and cloud platforms. However, conventional vision models operate with fixed compute and memory footprints for each input image or video, making them increasingly inadequate for dynamic runtime environments where edge resources vary (e.g., due to device contention) and cloud requests fluctuate (e.g., due to bursts in demand), even as latency requirements remain stringent. In this talk, I will present our recent work on adaptive inference for multimodal large language models in visual reasoning, where the execution of vision models can be dynamically reconfigured based on input content and available compute budgets. I will highlight how adaptive mechanisms can selectively reduce input representations and skip portions of network computation to explicitly trade accuracy for efficiency. Our overarching goal is to make vision models more efficient and robust when deployed beyond controlled settings, paving the way for intelligent systems that can gracefully adapt to the real world.

Creating effective short form videos that share your story is no easy task. While every day brings new technologies for video from generative video and audio models to multimodal large language models, using these technologies to invent new experiences and augment existing workflows remains challenging. We have been exploring how foundational language and vision models can make editing of short form narrative videos easier. In this talk I will describe how these powerful new technologies can be used in different stages of the video editing process from the very beginning when one is just getting started, to the middle when one is iterating, and the end when one wants to share many different versions targeting different audiences.

A 30-second ad in the 2011 SuperBowl cost $3 million. One particular ad, “The Force” by Volkswagen, earned millions of views in a few short days before it even aired on TV. In our work, we analyze what makes advertisements so powerful and important. In 2017, we proposed the first advertisement dataset, featuring over 64,000 image ads and 3,000 video ads. In 2018, we studied how video ads tell stories through narrative arcs culminating in a climax where a product enters the story to save the day. Fast-forward to 2025, we analyze what modern foundation models can achieve in terms of understanding and generation of creative advertisements. We discover that LLMs offer some advantages over MLLMs in terms of understanding ads, and can provide context to make generation of ads (with text-to-image models) more convincing. We also bridge an evaluation gap, namely in measuring how well ads follow their intended message, how persuasive and creative they are, and the nature of their inherent biases.