Welcome to the [AI Today] column! Here's your guide to exploring the world of AI every day, and every day we present you with hot content in the AI field, focusing on developers, helping you gain insight into technology trends and understand innovative AI product applications.
AI applications
Alibaba is building an AI e-commerce product, "Painted Frog".
[aibase synopsis:].The Peking University team launched the "Open SORA" AnimateDiff response to the SORA projectThe product is mainly aimed at merchants and talents, and its main functions are AI copywriting generation and AI image generation.
AI copywriting mainly realizes single-product grass planting, explosive article rewriting, clothing sharing, etc.
AI raw pictures will mainly train exclusive AI models for **, Tmall merchants and talents, and create the same product pictures of Internet celebrities.
[aibase synopsis:].Comfyui's transparent layer generation plugin LayerDiffusion is officially launchedThe Peking University team and Rabbit Exhibition launched the Replication SORA project, called Open SORA, and the framework has been built.
Open Sora plans to use a three-part framework that includes Video VQ-VAE, Denoising Diffusion Transformer, and Condition Encoder.
At present, the team needs more data and GPU for training, and Peking University alumni and AnimateDiff gods have responded positively.
Project Address:
[aibase synopsis:].comfyui-mana-nodes: A fun comfyui plugin with customizable font animationsNot only can transparent elements be generated directly, but also transparent elements that blend with the environment can be generated on top of existing ones.
The plugin mainly contains two nodes, one of which requires the comfyui-tooling-nodes plugin to be installed.
Currently, only the generation is supported, and the image quality is comparable to that of real commercial-grade transparent footage.
Project Address:
[aibase synopsis:].D-ID Agents: 1 ** + voice clone custom digital cloneFont animations with a high degree of customization are supported.
Users can customize the background color, font color, etc.
The plugin supports the use of local font files, providing more creative possibilities.
Project Address:
[aibase synopsis:].Zhejiang University & Microsoft launched the first-class editing framework Uniedit, which supports a variety of editing scenarios without training1 ** + clone your voice + synchronize the user knowledge base to customize your exclusive digital clone.
The digital human can carry out ** dialogue on your behalf with only 2 seconds delay, which can be used in ** meetings and other scenarios.
You can experience it by registering an account, and each user has 200 free access opportunities.
Experience address:
[aibase synopsis:].Dashtoon, an AI comic generation app, is a great tool for tweet makingUniedit excels in a wide range of editing scenarios.
Uniedit is unique in that it supports motion editing and a variety of appearance editing scenarios.
Uniedit makes use of temporal and spatial attention layers to achieve editing action and appearance.
Project Homepage:
[aibase synopsis:].GitHub has another AI tool DUST3R: 2 pictures and 3D reconstruction in 2 secondsComics can be generated in one go, providing ample room for editing and customization.
The consistency of the characters is mature and has high reference value.
Improve production efficiency and quality, and open up new possibilities.
Experience address:
[aibase synopsis:].What is worth buyingAI shopping assistant "small value" is online to provide shopping recommendations and suggestions across the networkDUST3R excels in monocular multi-view depth estimation and relative pose estimation tasks.
The authors' team took a new approach that allows for 3D reconstruction in any image without the need for camera calibration or a priori viewpoint pose information.
Dust3R achieves SOTA results in multiple tasks, demonstrating its power and applicability.
Project Entrance:
[aibase synopsis:].What's new in AIMeta AI proposes MobileLLM: a new way to deploy LLMs on mobile devicesThe self-developed AI shopping assistant "Xiaowei" was launched on the "What is worth buying" app.
Provide real-time *** and product word-of-mouth reviews to help users make informed shopping decisions.
Through the intent recognition model, we recommend high-quality products that meet the needs of users, bringing a new shopping experience.
[aibase synopsis:].Google Chrome adds more search features to the search barLLMs have challenges on mobile devices.
MobileLLM improves performance through deep and narrow structure design and parameter optimization.
Opens up new possibilities for applying LLMs in resource-constrained environments.
[aibase synopsis:].HKU Develops V-IRL Platform: Incorporating Real-World Maps into the Virtual Environment to Give AI Agents a Complete Life!Shopping-related searches in the Chrome app show thumbnails for a wider range of products and shopping categories.
Even with a poor network connection, Chrome provides search suggestions with "improved on-device features."
The search box displays suggested queries based on previous search records to provide users with a smarter, more personalized search experience.
[aibase synopsis:].Tsinghua University and Harbin Institute of Technology proposed the onebit method: the large model can be compressed to 1bit to maintain 83% performanceHKU and NYU's research team collaborated to develop the V-IRL platform, which integrates real-world maps and other information into the virtual environment to provide AI agents with a more real-life experience.
The V-IRL platform simulates the real environment, enabling agents to perform complex tasks and collaboratively solve problems across different tasks.
Based on V-IRL, the researchers conducted tests such as place recognition, visual question answering, and navigation, demonstrating the potential of AI for a wide range of real-world applications.
Project Address:
[aibase synopsis:].Tsinghua University and Harbin Institute of Technology jointly released **, which will try 1bit quantization, breaking through the limit of 2bit;
The new method combines 1-bit layer structure, SVID-based parameter initialization, and quantization-aware training.
1-bit quantization breaks through the 2-bit barrier and opens up new possibilities for efficiently running large models on mobile devices.
*Address: