AI-powered image generators have seen significant commodification over the past two years, with every major tech company trying to get a piece of the pie. Today, Meta has claimed a breakthrough in the field, announcing its state-of-the-art AI model, CM3Leon, which is said to surpass competitors in text-to-image generation.
Introducing CM3leon, a first-of-its-kind multimodal model that achieves state-of-the-art performance for text-to-image generation with 5x the compute efficiency of competitive models.
— Meta AI (@MetaAI) July 14, 2023
More details ➡️ https://t.co/VR12zkmLDs pic.twitter.com/jUnG7G1Fxf
CM3Leon also stands out as one of the first image generators capable of generating captions for images, a step towards more proficient image-understanding models. It utilizes an "attention" mechanism to evaluate the relevance of input data, which can enhance model training speed and parallelizability.
This new model is said to be more efficient than most transformers, requiring five times less compute and a smaller training dataset. The model was trained using a dataset of millions of licensed images from Shutterstock and contains over 7 billion parameters.
A key factor in CM3Leon’s superior performance is the use of supervised fine-tuning (SFT), a technique previously effective in training text-generating models. SFT improved CM3Leon’s performance in image generation and caption writing, enabling it to generate and edit images based on text instructions.
However, questions remain regarding bias, as generative AI models have been found to reinforce societal biases. Meta acknowledges that CM3Leon "can reflect any biases present in the training data," highlighting the need for transparency in the evolving AI industry. It remains unclear when or if Meta plans to release CM3Leon.
