Google said a new text to image AI of the company beat its competitors It is reported that the program, called imagen, receives text - such as "a picture of a Persian cat in a cowboy hat and red shirt playing guitar on the beach" - and outputs the results. Imagen can generate realistic images or artistic rendering.
Imagen is another achievement after other text image generators such as dall-e, vq-gan + clip and late diffusion models. Google said that when asked to compare images generated by imagen and other text image generators, its model was found to be superior to competitors in accuracy and image fidelity.
Google shares several examples of text prompts and images created by AI on its imagen website, but these may only represent the best results it produces. Google declined to comment on the incident.
Text image learning model shows the power of machine learning system. In this case, imagen eliminates the need to know how to use professional software such as Photoshop to create abstract images. As pointed out at the Google I / O conference earlier this month, AI systems are helping the company approach its vision for the future of environmental computing. Environmental computing refers to the day when people can intuitively use computers without specific system or code knowledge.
However, Google did not ignore the power of artificial intelligence from text to image, and the company chose not to release imagen to the public. Imagen searches the Internet for information to learn and create images. Because the Internet may be full of stereotypes and prejudices, these will eventually appear in imagen. Google said these biases include a preference for light skin color and some Western gender stereotypes. The company is also concerned that imagen may be used negatively by bad actors.
A white paper published by Google said: "generation methods can be used for malicious purposes - including harassment and the dissemination of misinformation, and raise many concerns about social and cultural exclusion and prejudice."
Google warned other AI manufacturers to be vigilant about publishing text image models to the public without paying close attention to the information that AI is trained.