an AI from Google that created very realistic photos from simple texts

Program As part of its google brainthe American giant presents Imagena project based on artificial intelligence able to illustrate a written text with an ultra-realistic photo (text to image). A certain number of details related to the results of the artificial intelligence have been revealed by the actors of the project and it must be admitted that the images generated by the AI ​​allow a stunning realism. Here is the essential to know about this revolutionary AI.

Imagen, the tool to transform your most beautiful ideas into images

To operate, Imagen takes into account all the details provided in writing. On this web page which tells us more about the program in question, several samples have been unveiled which prove the power ofImagen.

It is possible to see, among other things, extraordinary images paired with texts thanks to them were created. Even more interestingly, it should be noted that even the most surreal ideas are illustrated with great fidelity. Imagen is therefore a hub for those with an overflowing imagination.

A photo of a raccoon wearing an astronaut helmet, looking out the window at night. (Google credits)

A photo of a raccoon wearing an astronaut helmet, looking out the window at night. (Google Credits)

Thus, with this AI, simple writing is enough to create images comparable to those of specialized photo editing software. Also, no graphics skills are needed to use the AI.

You should know that Imagen takes into account all styles of images by demonstrating a very thorough understanding. That being said, it should be noted that Google does not deny having obtained some fuzzy results. On the other hand, the images that have been revealed on the website allow you to measure the scope of the program.

An art gallery exhibiting paintings by Monet. The art gallery is flooded. Robots circle the art gallery using paddle boards. (Google credits)

On this page intended to present the project, Google does not give all the details of the program. These are available on the arXiv platform. Google has hinted that the AI ​​uses language models to identify the idea hidden in a text.

Also, to reproduce the images with notable fidelity, broadcast models are used. Text transformation is managed by a large encoder T5-XXL as stated by the developer team. The created image, 64×64 px, then goes through two “Super Resolution” diffusion models to upsample the image, which goes to 256×256 and then to 1024×1024 pixels. A more efficient architecture has also been obtained to obtain a fast calculation speed. Named Efficient U-Netit optimizes memory management.

An AI that takes over its competitors

It should be noted that Google did not invent the concept text to image, but instead made it perform better. This means that programs of the same type were already available.

We denote among others: VQ-GAN+CLIP, Latent Diffusion Models and DALL-E 2 with which comparisons have even been made. The new Google AI would have obtained better results based on 2 evaluation modes.

To confirm this potency, another test was performed on Draw benchthen a last on Reference (which is based on human evaluations). Unsurprisingly, Imagen stood out as the most popular program on all counts.

Will Imagen be available to the general public?

A demo version limited to a few words is currently available on the Google site. But the general deployment and a publication of the source code are not currently on the agenda. Although the general public use of Image raises ethical problems in the face of a contrasting social situation.

In a society where discrimination and prejudice are still very present, such a program can contribute to accentuating these phenomena. Misuse is predictable and could be burdensome for marginalized minorities. Google admits that Imagen already works with some stereotypes.

To tell you the truth, the program has preferences for people with fair complexion and for the Western way of life. The image also takes into account all the data available on the web without filtering it, which could give more credence to fake news or violent messages. For all these relevant reasons, public use of Imagen is not yet on the agenda.

Leave a Comment