Li, X. (CSE) – Compute-Efficient Scaling of Fully-Open Visual Encoders
Vision encoders have demonstrated significant performance gains in visual generation and multimodal reasoning. These improvements are primarily attributed to the scaling of data, model capacity, and compute. However, this progress is becoming less accessible due to a lack of transparency in data curation and training recipes. In combination with the high compute requirements of foundation-scale pre-training, these factors hinder independent reproducibility.
In this dissertation, we democratize large-scale visual encoder training by developing compute-efficient, reproducible training recipes for video encoders, vision-language models (VLMs), and multimodal large language models (MLLMs). First, we challenge the common belief that scaling necessarily requires proportionally more resources. Specifically, we show that decoupled pre-training separates key factors such as space/time and token length, and learns strong priors first. This design yields dramatic efficiency gains across image, video, and generative modeling. Next, we address the challenge of undisclosed or inaccessible training data by releasing and systematically studying the curation of high-quality, large-scale datasets. We demonstrate that high-quality synthetic captions at scale enable vision-language models to learn stronger visual representations, especially when paired with training frameworks that unify contrastive and generative objectives. Lastly, building on these findings, we develop fully open vision encoders with complete training data, recipes, and checkpoints, and show that transparency can enable rather than hinder state-of-the-art performance as an MLLMs’ visual backbone.
Together, these contributions establish that openness and efficiency are mutually reinforcing, providing a reproducible foundation for the next generation of visual intelligence.
Event Host: Xianhang Li, Ph.D. Candidate, Computer Science and Engineering
Advisor: Cihang Xie
Zoom- https://ucsc.zoom.us/j/95801462664?pwd=koENnyV65jyPnkJYTbiYr1jaNsV5BE.1
Passcode- 782017
