You can now generate images using Google’s Imagen 3 AI model

Google on Thursday released its own artificial intelligence (AI) model for generating images, Imagen 3. The tech giant made no announcement prior to the release, and instead quietly released the model to users. Additionally, a research paper was also published in an online journal describing how the image generation model works. Currently, the text-to-image generation model is only available to users in the US, with no word on when it might roll out to users in other regions.

Imagen 3 AI model released by Google

The tech giant’s AI Test Kitchen is now offering users the opportunity to sign up for the platform and use the AI ​​model to generate images. The third generation of the Imagen model is said to gain improved texture generation and word recognition capabilities, as well as stricter prompt compliance.

Since the AI ​​model is only available in the US, Gadgets 360 was unable to test the platform. A Reddit user claimed that it was able to generate images in various styles, such as Nikon DSLR quality, GoPro style, wide-angle lens and more. However, the model reportedly struggled to generate close-up images with multiple people and underexposed images, which was possible with its predecessor.

Another area where Imagen 3 struggles is with limbs. The user claimed that the model produced erroneous results when using prompts such as “a man holding a cup of coffee.” The AI ​​would eventually generate additional limbs, create a random limb holding the object, or fuse the object and the limb together. The image generation model also reportedly has very strict censorship in prompts.

Google also published a study paper in the pre-print online journal arXiv. There, the company emphasized that it used a latent diffusion model, which is a variant of the diffusion model popularized by Stable Diffusion. The company also added that new methods were used to minimize the potential damage using the Imagen 3 model.

Interestingly, the free version of the Gemini chatbot can also generate images, but for this it uses the capabilities of Gemini. Imagen 3 is built on a different architecture and since the dataset largely contains images, it is better trained to generate AI images.