In the landscape of text-to-image models, the demand for high-quality visuals has surged. However, these models often need to grapple with resource-intensive training and slow inference, hindering their real-time applicability. In response, this paper introduces PIXART-δ, an advanced iteration that seamlessly integrates Latent Consistency Models (LCM) and a custom ControlNet module into the existing PIXART-α…

There’s a burgeoning interest in technologies that can transform textual descriptions into videos. This area, blending creativity with cutting-edge tech, is not just about generating static images from text but about animating these images to create coherent, lifelike videos. The quest for producing high-fidelity, aesthetically pleasing videos that accurately reflect the described scenarios presents a…

Developing large-scale datasets has been critical in computer vision and natural language processing. These datasets, rich in visual and textual information, are fundamental to developing algorithms capable of understanding and interpreting images. They serve as the backbone for enhancing machine learning models, particularly those tasked with deciphering the complex interplay between visual elements in images…

A pressing issue emerges in text-to-image (T2I) generation using reinforcement learning (RL) with quality rewards. Even though potential enhancement in image quality through reinforcement learning RL has been observed, the aggregation of multiple rewards can lead to over-optimization in certain metrics and degradation in others. Manual determination of optimal weights becomes a challenging task. This…

Researchers from Tel-Aviv University and Google Research introduced a new method of user-specific or personalized text-to-image conversion called Prompt-Aligned Personalization (PALP). Generating personalized images from text is a challenging task and requires the presence of diverse elements like specific location, style, or (/and) ambiance. Existing methods compromise personalization or prompt alignment. The most difficult challenge…

Combining CLIP and the Segment Anything Model (SAM) is a groundbreaking Vision Foundation Models (VFMs) approach. SAM performs superior segmentation tasks across diverse domains, while CLIP is renowned for its exceptional zero-shot recognition capabilities.  While SAM and CLIP offer significant advantages, they also come with inherent limitations in their original designs. SAM, for instance, cannot…

When the camera and the subject move about one another during the exposure, the result is a typical artifact known as motion blur. Computer vision tasks like autonomous driving, object segmentation, and scene analysis can negatively impact this effect, which blurs or stretches the image’s object contours, diminishing their clarity and detail. To create efficient…

Text-to-image synthesis is a revolutionary technology that converts textual descriptions into vivid visual content. This technology’s significance lies in its potential applications, ranging from artistic digital creation to practical design assistance across various sectors. However, a pressing challenge in this domain is creating models that balance high-quality image generation with computational efficiency, particularly for users…

The rapidly evolving domain of text-to-3D generative methods, the challenge of creating reliable and comprehensive evaluation metrics is paramount. Previous approaches have relied on specific criteria, such as how well a generated 3D object aligns with its textual description. However, these methods often must improve versatility and alignment with human judgment. The need for a…

The constant visibility of hands in our daily activities makes them crucial for a sense of self-embodiment. The problem is the need for a digital hand model that is photorealistic, personalized, and relightable. Photorealism ensures a realistic visual representation, personalization caters to individual differences, and reliability allows for a coherent appearance in diverse virtual environments,…