Grok Adopts Aurora for Enhanced Photorealistic Image Generation

Grok’s ability to generate images will now rely on Aurora, a new and independent image generator. Previously, the Flux model from Black Forest Labs was responsible for image generation in the chatbot. The current status of the partnership with the German company is unknown. Aurora seems to have as few restrictions as the entire Grok chatbot.

Elon Musk had Grok developed by his company xAI, founded in the summer of 2023. Grok is an AI chatbot available to paying users of X. It is known for having fewer rules and boundaries in text and image generation compared to other AI services. Musk describes this as “humorous interactions,” like when the image generator gives the Pokémon Pikachu a machine gun.

Grok provides photorealistic images and has multimodal capabilities. This approach has not changed, but the provider or AI model implementing these features has. xAI refers to an update for Grok, introducing an “autoregressive Mixture of Experts network.” Aurora has been trained on billions of images from the internet, offering a “deep understanding of the world.” A Mixture-of-Experts model consists of several specialized models acting as experts, with only the necessary ones responding to a request, improving performance without increasing computational demand. Autoregressive means the model predicts tokens differently from previous diffusion models that create images using noise.

With Aurora, Grok aims to produce photorealistic images and accurately execute prompts. The model is also multimodal, allowing users to input and edit their own images. Initial testers have used Aurora to generate images of people like Musk or Sam Altman, CEO of OpenAI, which appear very realistic.

The previous image generator in Grok was developed by Black Forest Labs. Their founders were formerly with Stability AI, known for the diffusion-based image generator Stable Diffusion. According to Black Forest Labs, Flux.1 features a hybrid architecture of multimodal and parallel diffusion-transformer blocks, claiming to surpass models like Midjourney v6.0, Dall-E 3, and SD3-Ultra in standard benchmarks. This raises the question of whether Aurora is even better or why xAI decided to switch providers.