Google released its in-house artificial intelligence (AI) model for image generation, Imagen 3, on Thursday. The tech giant did not make any announcement for the release, and instead released the model quietly to users. Additionally, a research paper detailing the workings of the image generation model was also published in an online journal. Currently, the text-to-image generation model is only available to users in the US, and there is no word on when it might be rolled out to users in other regions.
Imagen 3 AI Model Released by Google
The tech giant's AI Test Kitchen is now allowing users to sign up to the platform and use the AI model to generate images. The third generation of its Imagen model is said to get improved texture generation and word recognition capabilities as well as stricter prompt adherence.
Since the AI model is only available in the US, Gadgets 360 was not able to test out the platform. However, a Reddit user claimed that he was able to generate images in various styles such as Nikon DSLR quality, GoPro style, wide angle lens, and more. However, the model is said to be struggling with generating close-up images with multiple people and underlit images which was possible with its predecessor.
Another area where Imagen 3 struggles is limbs. The user claimed that the model was producing erroneous results when using prompts such as “a guy holding a cup of coffee”. The AI would end up generating extra limbs, creating a random limb holding the object, or fusing the object and the limb. The image generation model is also said to have very strict censorship in prompts.
Google also published a research paper in the pre-print online journal arXiv. There, the company highlighted that it used a latent diffusion model, which is a variant of the diffusion model popularised by Stable Diffusion. The company also added that new methods have been used to minimise the potential harm using the Imagen 3 model.
Notably, the free tier of the Gemini chatbot can also generate images, but it uses Gemini's capabilities for this. Imagen 3 is built on a different architecture and since its dataset largely contains images, it is better trained to generate AI images.
