xAI releases image generation model "Aurora" with "Grok". Is it ready to use? We proved it with experiments and compared recent high-quality image generation in Japanese!

On December 9, 2024, xAI (x.ai), part of the X Corp. group that operates the former Twitter and other services, updated Grok's features and integrated a new autoregressive image generation model codenamed "Aurora," available on the 𝕏 platform, into its conversational AI service, "Grok."

https://x.ai/blog/grok-image-generation-release

"Aurora" is described as an autoregressive Mixture-of-Experts network trained to predict the next token from data where text and image data are interleaved. Because the model was trained on billions of internet examples, it has a deep understanding of the world. As a result, it is not only excellent at photorealistic rendering and accurately following text instructions, but it is also designed with native support for multimodal input beyond text, allowing users to draw inspiration from or directly edit images they provide.

Grok's new features are now available on the 𝕏 platform in some countries and are expected to be rolled out to all users within a week.

Image Generation

Grok is now capable of generating high-quality images in several areas where other image generation models struggle. It can render accurate visual details of real-world entities, text, and logos, and create realistic human portraits.

<Sample images are introduced at the end of the article>

Entity Generation

Prompt: Cybertruck under an aurora

Although the Tesla "Cybertruck" was released only recently, Grok successfully generated it under an aurora. Google Deepmind's "Imagen 3" and Ideogram's "Ideogram 2.0" barely managed to generate a truck-like object, while "Flux.1 Pro" generated a conventional trailer, and "Dall-E 3" displayed a message saying it "cannot generate" it.

Artistic Text

Prompt: Stars in a galaxy spelling "Grok"

Similarly, Grok, Imagen 3, and Ideogram 2.0 were able to depict the stars, while Dall-E 3 and Flux.1 Pro produced illustrations.

Meme Generation

Prompt: Anime Pepe

"Pepe" is a global internet meme by Matt Furie.

I think this is a very abstract and difficult prompt. It seems it was difficult for Flux.1 Pro.

Realistic Portrait

Prompt: A man with wet rainbow hair and a neon blue raincoat standing in heavy rain on the city street

Celebrity

Prompt: Elon Musk wearing an xAI jacket with a USA flag in the background

This might be a difficult subject for Imagen 3 and DALL-E 3. The depiction of the beard in Grok is excellent.

Image editing

It seems that Grok's new image generation model can accept images as input. This feature is scheduled to be released to 𝕏 platform users in the future.

Prompt: Make the cat anime style

Prompt: Add a cowboy hat

Prompt: Make him blonde

Prompt: Make it snowy

Image generation experiment in Japanese by the AICU editorial department

It wouldn't be like AICU to write an article based only on the official release, so we will get our hands dirty and evaluate it. We will dare to provide prompts in Japanese.

Grok 2 (internal details unknown)

*Although it is labeled as 'Grok 2 new', it is difficult to determine at this time whether it is Aurora.

Prompt: I want a metallic image of 'MASTER GUIDE'

It is certainly wonderful that this quality can be generated in one go with Japanese. The 'GROK' logo appears in the bottom right.

I tried changing the prompt to English.

It feels more modern than a boiler plate. I will also try generating Japanese text.

Ideogram 2.0

Although there are many usage limits, Ideogram 2.0 can actually be said to be an excellent service that can be used for free.

Japanese prompt: I want a metallic image of 'MASTER GUIDE'

Imagen 3

Imagen 3 can be used from the Gemini interface.

https://gemini.google.com/

It is impressive that it can generate properly even in Japanese.
I will try using an English prompt.

Flux1.1 pro ultra

I am experimenting with Flux.1.1 pro ultra via Fal
https://fal.ai/models/fal-ai/flux-pro/v1.1-ultra

Prompt: Metallic 'MASTER GUIDE' boiler plate (nameplate)

Prompt: metallic boiler plate of "MASTER GUIDE"

While it manages expressions like rust, the representation of the font itself might be starting to look inferior compared to other models.

Dall-E 3

I will try generating it with ChatGPT 4o (paid version).

Prompt: 'Create an image: Metallic 'MASTER GUIDE' boiler plate (nameplate)'

English seems to be more reliable.

If you look closely, there are mistakes in the text.

Comprehensive evaluation as of December 2024

I have expressed the above experiment in a scorecard, adding 'text rendering ability,' 'Japanese interpretation,' and 'authenticity information.'

Since it is difficult to evaluate differences in entity generation ability, expressive power, or style representation such as anime, I evaluated it based on text rendering ability and Japanese interpretation as a Japanese user (there are no models among these that can render Japanese text). Therefore, in the comprehensive evaluation, Grok2 (Aurora) certainly receives a high rating. In particular, while there are models like 'Flux1.1 pro ultra' and Stability AI's 'Stable Diffusion 3.5,' I evaluated the other models from the perspective of being 'available for free,' specifically for hobbyists or casual users who want to use image generation interactively in a text chat.

Actually, xAI has a partnership with black forest labs., the developer of Flux, and announced 'FLUX.1 now on Grok 2.0' on August 14th.

In other words, while Grok2's capabilities until now (and currently) may have been provided by Flux from Black Forest Labs (BFL), the newly released 'Aurora' is referred to as 'a new autoregressive image generation model,' so it is possible that an improved version of Flux is being used internally.

In terms of whether the current Grok2 is using Aurora, based on the experiment above, since Flux has significantly inferior Japanese interpretation capabilities, it can be said that we have confirmed that 'the current Grok2 has become Aurora'.

While there are some unclear aspects regarding BFL's business model, depending on their future relationship with xAI and the deployment of Aurora, it is possible that they will split into professional use and consumer use via B2B channels like xAI. Is this part of the 'war without honor and humanity' between Big Tech companies?

Only DALL-E 3 and Imagen 2 have authenticity information

In parallel with this investigation, I checked for authenticity using C2PA.
I checked whether the downloaded images contained embedded information.
https://contentcredentials.org/verify

In reality, it seems that DALL-E 3 is the only model that embeds authenticity information using C2PA, in which Adobe, OpenAI, Stability AI, Microsoft, AWS, and Google are participating.

Imagen could be verified using 'SynthID', an authenticity verification technology (image synthesis detection technology) developed by Google DeepMind. If you select 'About this image' using the 'Google Lens' smartphone app, it will display 'Generated by Google AI'.

※Thank you, Tadashi Honjo, for pointing this out.

Furthermore, if you want to perform comprehensive image generation detection, the 'AI-generated content detection tool' provided by Hive is recommended.

It comprehensively determines whether content is AI-generated and infers which model was used.

That concludes my summary of the battle between Big Tech companies in text-to-image generation, which is fitting for a wrap-up of 2024.

Also, at the same time, it seems that OpenAI has finally released 'Sora,' a video generation model that acts as a world simulator.

https://openai.com/sora/

I look forward to new models from Stability AI and the entry of players from China in the future!

If you want to learn the Stability AI API, this is recommended!
'Super Introduction: Image Generation AI with Stable Diffusion'
Generative AI Creative AICU
https://j.aicu.ai/coloso1

Aurora Gallery

Finally, I will introduce some generation examples from the official gallery.

https://x.ai/blog/grok-image-generation-release