Join our in-person meetup to hear talks from experts on cutting-edge topics across AI, ML, and computer vision.

Some comments on predicting what people like to hear from how it sounds.

Generative AI has made remarkable strides in producing photorealistic images, videos, and multimodal content. Yet, aligning these generations with human users, while ensuring spatial coherence, logical consistency, and deployment scalability, remains a major challenge, especially for real-world media platforms. In this talk, I will present our recent progress in enhancing the reasoning and control capabilities of image/video generative models, structured around four key pillars:

1. Efficiency & Scalability — with systems like ECLIPSE and FlowChef;
2. Control & Editing — including Lambda-ECLIPSE and RefEdit;
3. Reliability & Security — through efforts such as SPRIGHT, REVISION, R.A.C.E., and WOUAF;
4. Evaluation & Metrics — via benchmarks/metrics like VISOR, ConceptBed, TextInVision, and VOILA.

Together, these contributions outline a cohesive vision for building controllable, robust, and scalable generative systems, core to advancing personalization, understanding, and automated content workflows in media streaming and beyond.

Diffusion models have achieved impressive performance in generating high-quality and diverse synthetic data. However, their success typically assumes a class-balanced training distribution. In real-world settings, multi-class data often follow a long-tailed distribution, where standard diffusion models struggle -- producing low-diversity and lower-quality samples for tail classes. While this degradation is well-documented, its underlying cause remains poorly understood. We investigate the behavior of diffusion models trained on long-tailed datasets and identify a key issue: the latent representations (from the bottleneck layer of the U-Net) for tail class subspaces exhibit significant overlap with those of head classes, leading to feature borrowing and poor generation quality. Importantly, we show that this is not merely due to limited data per class, but that the relative class imbalance significantly contributes to this phenomenon. To address this, we propose COntrastive Regularization for Aligning Latents (CORAL), a contrastive latent alignment framework that leverages supervised contrastive losses to encourage well-separated latent class representations. We highlight the effect of CORAL for many large imbalanced datasets.

This talk shows how researchers can use FiftyOne to make their work more visible and impactful. We will cover practical ways to showcase datasets, models, and results, connect research tools, and share work with a broader community. The goal is to help researchers extend the reach of their research beyond papers and into real, reusable workflows.