TOWARDS UNSUPERVISED BLIND FACE RESTORATION USING DIFFUSION MODELS

An image restoration model is trained to fix damaged or unclear images, especially faces. First, it learns from a set of synthetic images to get a basic understanding. Then, real damaged images are fed into this model to create initial restored versions. These initial images are further processed using a special technique called denoising diffusion to create better target images. Finally, the model is improved by adjusting its settings based on how well it performed in restoring the images.

ADAPTIVE FOVEATED ENCODER AND GLOBAL MOTION PREDICTOR

An adaptive video encoder can adjust how it compresses video based on what a viewer is focusing on and how their headset is moving. It gathers information about the headset to understand where the viewer's attention is directed. Using this information, it sets specific rules for encoding video. This means it can focus on important parts of the image while saving data on less important areas. The result is a more efficient video experience that looks good without using too much data.

VISUAL CHAIN-OF-THOUGHT REASONING FOR MULTIMODAL LANGUAGE MODELS

A multimodal assistant system can understand both images and written tasks. It takes an image and a description of what needs to be done with that image. The system creates a visual Chain-of-Thought (v-CoT) prompt that outlines steps to follow in order to complete the task. These steps involve describing important details from the image and thinking through them to find a solution. Finally, the system provides the answer based on this reasoning process.

NAMESPACES FOR PROTECTED MULTI-TENANT PACKET PROCESSING PIPELINES

A new system allows different users to process data packets safely in a shared network. It uses separate areas called namespaces to keep each user's programming isolated from others. This isolation helps protect the data processing pipeline, which can serve multiple users at once. Even though the programming happens in these separate areas, it is turned into a single set of instructions that runs efficiently in the network. The goal is to improve how quickly data packets move through the system while reducing delays.

COUNTERFACTUALS WITH FEATURE PREFERENCES FOR CONSISTENT AND DIVERSE EXPLANATIONS

An artificial intelligence system can identify specific features of a machine learning model. When it detects a feature, it uses a special tool called a Counterfactual Engine to create a new scenario by changing that feature. This involves adjusting the importance of the feature and making some changes to it. The Counterfactual Engine then takes this new scenario and processes it further. The goal is to provide clear and varied explanations for how changes in features affect outcomes.

CUSTOM IMAGE AND CONCEPT COMBINER USING DIFFUSION MODELS

A new method allows for creating images by combining different types of inputs, like multiple pictures and text. First, the system takes these inputs and transforms them into a format that a computer can understand. Then, it uses a machine learning model to blend the information from the images and text. The result is a new image that includes elements from the original pictures. This technology can help in generating unique visuals based on specific ideas or themes.

Context-Aware Speech Recognition Using Prompts for Language Learners

A new method helps language learners improve their speaking skills. It starts by creating a prompt that asks the user to respond in a specific language, along with an example of the correct answer. This prompt is sent to the user's device, where they can record their spoken response. After receiving the audio, the system analyzes it using a trained speech model. Finally, it provides feedback on how well the user spoke in the target language.

NOVEL SYSTEM AND METHOD ADAPTED TO ENABLE SIMULTANEOUS PLAY OF ONE OR MORE SPORTS GAMES IN THE SAME FACILITY WITH AUTOMATED BALL DELIVERY, TRACKING AND COLLECTION

A new system allows multiple sports games to be played at the same time in one place. It features an automated way to deliver and track balls, making it especially useful for baseball and softball. The system includes a special machine that tosses balls up from the ground and devices that track players and balls. It also has a collection system that removes the need for fielders, creating a unique batting range experience. Additionally, it offers interactive screens and augmented reality to help players communicate and play together, even if they are not in the same location.

System and Method for Markerless Motion Capture

A computer system uses a camera to capture a person's movements without needing any special markers. It then applies these movements to a 3D model of an avatar that represents the user, which is linked to a unique digital token. This token helps identify the user in both traditional and new online spaces. The system creates a video that shows the avatar moving in real-time, mimicking the user's actions like hand gestures and facial expressions. Finally, it generates a digital token that proves ownership of this video.

SUPPLY CHAIN OPTIMIZATION WITH REINFORCEMENT LEARNING

A new system helps improve how products move through a supply chain. It uses a type of artificial intelligence called reinforcement learning to make better decisions. This method focuses on managing multiple levels of distribution, ensuring that products reach their destinations efficiently. By learning from past experiences, the system can adapt and find the best ways to optimize the supply chain. Overall, it aims to make the process smoother and more effective.