.png)
December 13, 2024
The AI chatbot Grok from xAI, available for paying X users, has received an upgrade to its image generator. Instead of the previous model Flux from Black Forest Labs, Aurora, a proprietary model developed by xAI, is now being used. The future of the collaboration with Black Forest Labs is currently unclear.
Aurora is based on an autoregressive Mixture-of-Experts network and was trained with billions of images from the internet. This allows the model to develop a comprehensive understanding of the world and generate photorealistic images. In contrast to diffusion models, which generate images through noise, Aurora predicts the next token in a sequence. This autoregressive process is supposed to lead to higher accuracy in the implementation of text inputs (prompts).
The Mixture-of-Experts model consists of specialized sub-models, each acting as experts for specific tasks. When a request is made, only the relevant experts are activated, which optimizes performance while maintaining the same computational effort. In addition to text input, Aurora also supports multimodal input, allowing, for example, users to upload and edit their own images. A corresponding editing function for X is planned.
Initial tests with Aurora show promising results. Generated images of people like Elon Musk or Sam Altman appear deceptively real. The realistic representation of people, texts, logos, and details, which often pose difficulties for other image generators, is one of Aurora's strengths.
Grok, which is characterized by fewer restrictions on content generation, is expected to become even more powerful with Aurora. While critics complain about the lack of guardrails in Grok, Elon Musk sees it as "humorous interactions," such as depicting Pikachu with a machine gun.
The reasons for the switch from Flux to Aurora are not officially known. Flux.1, the previous model from Black Forest Labs, was based on a hybrid architecture of multimodal and parallel diffusion-transformer blocks. It is said to have already surpassed models like Midjourney v6.0, Dall-E 3, and SD3-Ultra in benchmarks. Whether Aurora now further increases this performance or whether other reasons were decisive for the switch remains open.
xAI emphasizes the advances that Aurora enables in the field of multimodal image generation. The company is currently actively looking for new employees to work on this technology.
Sources:
