HomeAI News
Meta AI open source ImageBind! Let AI evolve again

Meta AI open source ImageBind! Let AI evolve again

Hayo News
Hayo News
May 10th, 2023
View OriginalTranslated by Google

MetaAI is a young man with many broken bloods in the Metaverse and Web 3.0 fields. Now he wields the open source sword and kills all directions, leading the development of the AIGC field.

In the past few months, MetaAI has open sourced many practical projects on GitHub, such as Segment Anything (SAM), which can automatically segment all items in a picture or video, making image editing easier; DINOv2, obtained through self-supervision Visual features can directly promote the progress of computer vision technology without fine-tuning; Animated Drawings uses AI capabilities to quickly add animation effects to paintings and so on. There are too many such useful items to list. Today, Meta shines again, announcing the open source ImageBind, so that the model can communicate across 6 different modalities (image, text, audio, depth, temperature and IMU data)! Here is a video posted by Zuckerberg on Facebook a few days ago, so that you can experience the powerful capabilities of ImageBind for yourself:

GitHub: https://github.com/facebookresearch/ImageBind

In order for the AI model to achieve capabilities closer to humans, it is necessary to enhance the support capabilities of multiple modalities. The reason why we can see the bustling streets, hear the whistle on the road, and feel the heat wave in the hot summer is because we have the sensory abilities of hearing, smell, taste and vision, so that we can better communicate with this world. world to interact. In order to make AI's capabilities closer to humans, we need to give it more capabilities so that it can perceive the world more accurately. Previously, to achieve association retrieval between different modalities, we needed to train multiple data sets at the same time. But now with ImageBind, we can generate images directly from audio. For example, when we let AI listen to the sound of ocean waves, it can directly generate an image of the sea, thus saving training time and cost. From the interface point of view, the current AI is like a human being, able to imagine the corresponding picture based on the sound. In addition, ImageBind also has built-in 3D perception and IMU sensors for measuring data such as physical motion, speed, and rotation, allowing AI to experience changes in our physical world more immersively. In addition, ImageBind also provides a new type of memory retrieval method, which can directly use the combined data of text, audio and images to search and match pictures, videos, audio files or text messages, so that the previous AI applications Generate higher quality content. For example, applying ImageBind to the field of video editing, AI can search and match higher-quality video clips based on the sound, image, text and other information we provide, realizing the function of one-click video editing. In traditional AI systems, each modality has specific embeddings, making it difficult to interact and retrieve between different modalities, and we cannot accurately retrieve related images and videos through audio. However, ImageBind does just that. It enables cross-modal retrieval by aligning the embeddings of six different modalities into a common space.

As a multimodal model, ImageBind integrates the SAM and DINOv2 I mentioned earlier, thus further enhancing its own capabilities. ImageBind binds various modalities together to build a bridge for seamless communication, which is also its core function. In the Make-A-Scene tool developed by MetaAI, we can generate images from text. And now, with the help of ImageBind, we can generate images directly from sound. This allows AI to gain a deeper understanding of human emotions and understand our emotions, thereby providing better services. In addition, the cross-modal communication capability of ImageBind will also bring mutual promotion between different modalities. The enhanced capabilities of each modality will drive the progress of the other modality, creating a situation similar to a snowball effect. In order to verify this, the MetaAI technical team conducted a benchmark test and found that ImageBind is significantly better than other professional models in terms of audio and depth, which also benefits from the experience absorbed and summarized by AI from other modalities.

In the foreseeable future, video editing will become easier and simpler. For example, when you hold up your mobile phone and record a video of a seaside sunset, AI can automatically generate copywriting and subtitles based on the video content, and match it with appropriate background music. Even AI may be able to directly generate a video MV for the singer through a song. In VR and AR games, users can interact with game characters through various voices, gestures, and head movements to enhance the interactivity and immersion of the game. In the medical field, doctors can collect patient's condition information through various methods such as voice and images, and then process and analyze it through machine learning and other technologies to more accurately obtain diagnosis results and treatment plans. Although currently ImageBind only covers six modalities, with the integration of more sensory functions (such as smell, touch, etc.), the ability of the AI model will become more powerful, thus bringing more interesting and practical AIGC projects and application scenarios. The emergence of this project will bring a wider range of application scenarios for AIGC technology, and more interesting and practical AI projects and application scenarios will be ushered in in the future. We are one step closer to the arrival of general artificial intelligence.


no dataCoffee time! Feel free to comment