Meta introduced the new CM3Leon AI model with the best performance

During the last two years, artificial intelligence models of image production have become very popular, and the technical barriers to their use have almost been removed. Of course, this does not mean that the performance of these models has become stable and reached a suitable level; But Meta claims that with the new model CM3Leon He has made an important achievement and made the best example in this field.

Meta today unveiled a new artificial intelligence model called CM3Leon (aka Chameleon or “sunworshipper”) and says that this system can convert text into images. the best Demonstrate performance. This model is also one of the first models that can explain or caption to write

CM3Leon is a convertible model orTransformer” is that using a mechanism called “Attention(Attention) tries to evaluate the relationship of existing data such as text or images to each other. This attention, along with other architectural features in transformer models, can increase the speed of training models. Making larger transformer models is possible by increasing the processing power.

The required power for CM3Leon Meta is five times less than similar models

Meta claims that the CM3Leon is more optimized than most Transformer models, as it five times processing power fewer It needs and can be trained with a smaller database of data. The company used millions of images from Shutterstock to train its model.

The most powerful version of CM3Leon ca 7 billion parameters which is almost twice the DALL-E 2 model from OpenAI company. One of the key features of Meta’s new artificial intelligence model, which has helped to increase its power, is a technique called “Supervisory optimization(SFT) is. This technique was already used in text models like ChatGPT and now it has proven itself in image models.

The CM3Leon Meta model can also understand commands related to editing existing images. In addition, as we said, this model is able to write a short or long description for the images. The performance of CM3Leon in this field is even better than models for creating captions for images such as Flamingo and OpenFlamingo.

Meta has not yet announced whether or not it plans to launch the CM3Leon AI model, and if so, when it will happen.