Grok Image Generation Release

xAI has introduced Aurora, a new autoregressive image generation model integrated into Grok. This update enables Grok to produce high-quality images with a focus on photorealistic rendering and strict adherence to text instructions, while introducing native multimodal support for image-to-image editing.

Aurora Model Architecture and Training

Aurora is built as an autoregressive mixture-of-experts (MoE) network. The model is designed to predict the next token from data that interleaves text and images. To achieve a deep understanding of the world and high fidelity in rendering, xAI trained Aurora on billions of examples sourced from the internet.

Core Image Generation Capabilities

Aurora allows Grok to generate high-quality images across several domains where traditional image generation models often struggle. Key capabilities include:

• Real-World Entity Rendering: Precise visual details of real-world objects and entities.
• Text and Logo Integration: The ability to render accurate text and logos within images.
• Human Portraits: Creation of realistic portraits of humans and celebrities.
• Diverse Content Types: Support for a wide range of styles, from photorealistic landscapes and cyberpunk cities to artistic sketches and comic-style illustrations.

Multimodal Input and Image Editing

Beyond text-to-image generation, Aurora features native support for multimodal input. This allows the model to use user-provided images as a reference for inspiration or as a direct target for editing. For example, users can provide an image of a cat and prompt the model to "Make the cat anime style," resulting in a stylized version of the original image.

Availability and Deployment

As of December 9, 2024, these image generation capabilities are available on the  platform in select countries. xAI stated that the feature will roll out to all users within one week of the announcement.