Video generation models have become a core technology for creating dynamic content by transforming text prompts into high-quality video sequences. Diffusion models, in particular, have established themselves as a leading approach for this task. These models work by starting from random noise and iteratively refining it into realistic video frames. Text-to-video (T2V) models extend this…

The core idea of Multimodal Large Language Models (MLLMs) is to create models that can combine the richness of visual content with the logic of language. However, despite advances in this field, many models struggle to connect the two domains effectively, leading to limited performance in complex reasoning tasks that involve visual components. A major…

Manipulating lighting conditions in images post-capture is challenging. Traditional approaches rely on 3D graphics methods that reconstruct scene geometry and properties from multiple captures before simulating new lighting using physical illumination models. Though these techniques provide explicit control over light sources, recovering accurate 3D models from single images remains a problem that frequently results in…

LLMs have made significant strides in language-related tasks such as conversational AI, reasoning, and code generation. However, human communication extends beyond text, often incorporating visual elements to enhance understanding. To create a truly versatile AI, models need the ability to process and generate text and visual information simultaneously. Training such unified vision-language models from scratch…

Text-to-image (T2I) generation has evolved to include subject-driven approaches, which enhance standard T2I models by incorporating reference images alongside text prompts. This advancement allows for more precise subject representation in generated images. Despite the promising applications, subject-driven T2I generation faces a significant challenge of lacking reliable automatic evaluation methods. Current metrics focus either on text-prompt…

The CLIP framework has become foundational in multimodal representation learning, particularly for tasks such as image-text retrieval. However, it faces several limitations: a strict 77-token cap on text input, a dual-encoder design that separates image and text processing, and a limited compositional understanding that resembles bag-of-words models. These issues hinder its effectiveness in capturing nuanced,…

Despite rapid advances in vision-language modeling, much of the progress in this field has been shaped by models trained on proprietary datasets, often relying on distillation from closed-source systems. This reliance creates barriers to scientific transparency and reproducibility, particularly for tasks involving fine-grained image and video understanding. Benchmark performance may reflect the training data and…

3D self-supervised learning (SSL) has faced persistent challenges in developing semantically meaningful point representations suitable for diverse applications with minimal supervision. Despite substantial progress in image-based SSL, existing point cloud SSL methods have largely been limited due to the issue known as the “geometric shortcut,” where models excessively rely on low-level geometric features like surface…

Recent advancements in AI scaling laws have shifted from merely increasing model size and training data to optimizing inference-time computation. This approach, exemplified by models like OpenAI o1 and DeepSeek R1, enhances model performance by leveraging additional computational resources during inference. Test-time budget forcing has emerged as an efficient technique in LLMs, enabling improved performance…