AI generates great photos using text-to-image

In January 2021, I made Open AI Artificial Intelligence Research Group with Dall-E . Project talk about yourself. Here, AI is rendering concepts visually. Users start with a current image, select a region and tell the AI ​​that this section should be addressed. This should make photo editing easier. Alternatively, image search and creation of image variations can be trained. Now Dall-E is facing competition from Google’s “Imagen”. In the meantime, the company presented the best results to the public.

Imagen shouldn’t be published – AI needs human help

For all images created with Imagen, the text under the image was the specific vector for the program. For example, if you enter “robot couple feeding with the Eiffel Tower in the background,” you will receive an automatically generated image of a robot couple enjoying dinner in front of the Eiffel Tower. According to Google, Imagen is based on the new Drawbench standard.

Drawbench is a list of about 200 text prompts that the Google team has entered into the text-to-image generators. Each result was then evaluated by human critics and then fed to artificial intelligence. According to Google, Imagen should not be published.

Instead, this technology should be used to train algorithms that can recognize fake news. In order to achieve good results, the algorithms should be fed as many training data sets as possible and an evaluation should be made of how the program handles them. According to Google researchers, many existing data sets tend to reflect stereotypes or other socially undesirable associations. AI software is reaching its limits when evaluating data and needs human help here.

Also worth reading: Nvidia: GPU-based AI can develop next-generation chips

Also worth reading: Amazon is getting rid of Alexa: Fire TV Stick 4K, Echo Show, Echo Dot, Ring, and Fire Kids for Very Cheap

Google Imagen Text-to-Image Generator Set:

• Google introduced a new text-to-image generator Imagen, which uses the Drawbench scale. In the meantime, the company presented the best results to the public.
• After entering text, Imagen can automatically create an image. According to Google, the generator should not be published. Instead, this technology should be used to train algorithms that can recognize fake news.
• Google researchers caution that many training data sets for algorithms tend to be modular. Thus, AI programs need human assistance in evaluating their data. Google, for example, manually evaluated about 200 text prompts on Drawbench and used them to train Imagen.

Sources: The GoogleAnd the Take Crunch