Google has introduced the Imagen neural network, better than DALL-E 2
The Google Brain team has introduced a neural network artist who recognizes text using large language models and then produces a photorealistic image. Imagen generates the first image of 64×64 pixels, and then increases the resolution to 256×256 and 1024×1024 pixels, adding details in the process. AI Imagen has already surpassed DALL-E 2 in image quality.
“Our key discovery is that universal large language models (such as T5) pre-trained on text arrays are surprisingly effective at encoding text for image synthesis: increasing the size of the Imagen language model significantly improves both sampling accuracy and image alignment and text“, – said the developers.
The Google team also reported ethical issues, although they did not provide details on the disturbing content generated by Imagen. However, Google Brain noted that the neural network ‘encodes several social prejudices and stereotypes, including a general bias towards creating images of people with lighter skin tones and a penchant for images depicting different professions that conform to Western gender stereotypes.’
All because Imagen ‘fed’ data sets from the Internet without 100% pre-selection. And information of this kind often shows ‘social stereotypes, repressive views and degrading or otherwise harmful associations with marginalized identity groups.’ Much of the training data was filtered out of unwanted content. But they also used the LAION-400M dataset, which contains a wide range of unacceptable content, pornographic images, images of racists and harmful social stereotypes.
Therefore, the source code is not presented to the public.
