VLIC: Vision-Language Models As Perceptual Judges for Human-Aligned Image Compression

A novel approach called Knowledge Transfer (KT) has been introduced to enhance Vision-Language Models (VLMs) by enabling them to learn new visual concepts solely from textual descriptions. This method aligns visual features with text representations, allowing VLMs to visualize previously unknown concepts without relying on visual examples or external generative models.

Vision-language models (VLMs) have been found to struggle with 3D-related tasks, which are essential for applications in robotics and embodied agents. To address this issue, a new framework called SandboxVLM has been introduced, which utilizes abstract bounding boxes to enhance the encoding of geometric structures and physical kinematics, thereby improving spatial intelligence in VLMs.

PerTouch has introduced a unified diffusion-based framework for image retouching that enhances visual quality while aligning with user preferences. This method employs parameter maps to facilitate fine-grained adjustments and incorporates mechanisms for improved semantic boundary perception, allowing for a more personalized retouching experience.

The introduction of the Temporal Understanding in Autonomous Driving (TAD) benchmark addresses the significant challenge of temporal reasoning in autonomous driving, specifically focusing on ego-centric footage. This benchmark evaluates Vision-Language Models (VLMs) through nearly 6,000 question-answer pairs across seven tasks, highlighting the limitations of current state-of-the-art models in accurately capturing dynamic relationships in driving scenarios.