Meta, the company behind the popular social media platform, has unveiled its latest AI image generation model called CM3leon (pronounced chameleon). This new model promises to be more efficient and effective in creating images from text and generating captions for existing images.
While AI-generated images are not new, Meta’s approach to building CM3leon sets it apart. Instead of using diffusion models like other image generation tools, CM3leon utilizes a token-based autoregressive model. This token-based approach, although more computationally expensive, has shown to produce superior results in terms of global image coherence.
One remarkable achievement of CM3leon is its efficiency. Meta researchers have demonstrated that the token-based autoregressive model can outperform diffusion models while requiring five times less computational resources. This breakthrough brings text-to-image generation models to a new level.
The process behind CM3leon’s development starts with a retrieval-augmented pre-training stage. In contrast to scraping publicly available images, which has raised legal concerns, Meta uses licensed images from Shutterstock, ensuring image ownership and attribution are respected without compromising performance.
Following the pre-training stage, CM3leon undergoes a supervised fine-tuning (SFT) process that optimizes both resource utilization and image quality. This approach, similar to the one used by OpenAI in training ChatGPT, enhances the model’s ability to understand complex prompts, making it highly effective in generative tasks.
The sample sets of generated images shared by Meta showcase CM3leon’s impressive capabilities. The model demonstrates a deep understanding of complex, multi-stage prompts and produces extremely high resolution images as a result.
Although CM3leon is currently a research effort, it is highly likely that Meta will make this technology available in the future, given its power and efficiency in image generation. However, no official timeline or platform for its release has been announced.
In conclusion, Meta’s latest AI image generation model, CM3leon, stands out for its efficiency and performance. By leveraging token-based autoregressive models, CM3leon achieves state-of-the-art results while requiring fewer computational resources. While its availability as a public service is still uncertain, CM3leon’s potential impact on generative AI is unquestionable.
