Video generation has rapidly become a focal point in artificial intelligence research, especially in generating temporally consistent, high-fidelity videos. This area involves creating video sequences that maintain visual coherence across frames and preserve details over time. Machine learning models, particularly diffusion transformers (DiTs), have emerged as powerful tools for these tasks, surpassing previous methods like…

Understanding and analyzing long videos has been a significant challenge in AI, primarily due to the vast amount of data and computational resources required. Traditional Multimodal Large Language Models (MLLMs) struggle to process extensive video content because of limited context length. This challenge is especially evident with hour-long videos, which need hundreds of thousands of…

Long Video Segmentation involves breaking down a video into certain parts to analyze complex processes like motion, occlusions, and varying light conditions. It has various applications in autonomous driving, surveillance, and video editing. It is challenging yet critical to accurately segment objects in long video sequences. The difficulty lies in handling extensive memory requirements and…

The rapid progress of text-to-image (T2I) diffusion models has made it possible to generate highly detailed and accurate images from text inputs. However, as the length of the input text increases, current encoding methods, such as CLIP (Contrastive Language-Image Pretraining), encounter various limitations. These methods struggle to capture the full complexity of long text descriptions,…

Large Language Models (LLMs) have demonstrated remarkable progress in natural language processing tasks, inspiring researchers to explore similar approaches for text-to-image synthesis. At the same time, diffusion models have become the dominant approach in visual generation. However, the operational differences between the two approaches present a significant challenge in developing a unified methodology for language…

Parameter-efficient fine-tuning (PEFT) methods, like low-rank adaptation (LoRA), allow large pre-trained foundation models to be adapted to downstream tasks using a small percentage (0.1%-10%) of the original trainable weights. A less explored area of PEFT is extending the pre-training phase without supervised labels—specifically, adapting foundation models to new domains using efficient self-supervised pre-training. While traditional…

Dense geometry prediction in computer vision involves estimating properties like depth and surface normals for each pixel in an image. Accurate geometry prediction is critical for applications such as robotics, autonomous driving, and augmented reality, but current methods often require extensive training on labeled datasets and struggle to generalize across diverse tasks. Existing methods for…

Biomedical vision models are increasingly used in clinical settings, but a significant challenge is their inability to generalize effectively due to dataset shifts—discrepancies between training data and real-world scenarios. These shifts arise from differences in image acquisition, changes in disease manifestations, and population variance. As a result, models trained on limited or biased datasets often…

Large language and vision models (LLVMs) face a critical challenge in balancing performance improvements with computational efficiency. As models grow in size, reaching up to 80B parameters, they deliver impressive results but require massive hardware resources for training and inference. This issue becomes even more pressing for real-time applications, such as augmented reality (AR), where…

Artificial intelligence has significantly enhanced complex reasoning tasks, particularly in specialized domains such as mathematics. Large Language Models (LLMs) have gained attention for their ability to process large datasets and solve intricate problems. The mathematical reasoning capabilities of these models have vastly improved over the years. This progress has been driven by advancements in training…