Pre-trained vision models have been foundational to modern-day computer vision advances across various domains, such as image classification, object detection, and image segmentation. There is a rather massive amount of data inflow, creating dynamic data environments that require a continual learning process for our models. New regulations for data privacy require specific information to be…

Generative models have revolutionized fields like language, vision, and biology through their ability to learn and sample from complex data distributions. While these models benefit from scaling up during training through increased data, computational resources, and model sizes, their inference-time scaling capabilities face significant challenges. Specifically, diffusion models, which excel in generating continuous data like…

Generative Large Multimodal Models (LMMs), such as LLaVA and Qwen-VL, excel in vision-language (VL) tasks like image captioning and visual question answering (VQA). However, these models face challenges when applied to foundational discriminative VL tasks, such as image classification or multiple-choice VQA, which require discrete label predictions. The primary obstacle is the difficulty in extracting…

One of the major hurdles in AI-driven image modeling is the inability to account for the diversity in image content complexity effectively. The tokenization methods so far used are static compression ratios where all images are treated equally, and the complexities of images are not considered. Due to this reason, complex images get over-compressed and…

Latent diffusion models are advanced techniques for generating high-resolution images by compressing visual data into a latent space using visual tokenizers. These tokenizers reduce computational demands while retaining essential details. However, such models suffer from a critical challenge: increasing the dimensions of the token feature increases reconstruction quality but decreases image generation quality. It thus…