Artificial Intelligence

Parti, Google's new artificial intelligence to create images

While it is true that artificial intelligence is marking a before and after in any field that is being developed, the art world was not so promising until the arrival in the same year of three systems that are making people talk, now it is the turn of Parti.

Parti (Pathways Autoregressive Text-to-Image) is the alternative that many were waiting for, a system capable of generating photorealistic images based on an autoregressive model that allows more extensive text inputs and can make complex compositions, unlike DALL-E and its variants that use a model of image generation from diffusion text.

Parti works text-to-image generation as a sequence-to-sequence modeling problem, analogous to machine translation, so it could benefit from advances in large linguistic models, especially capabilities unlocked by scaling data and model size. The target results are sequences of image tokens rather than text tokens in another language.

By using the powerful image tokenizer, ViT-VQGAN, to encode images as discrete token sequences, it leverages its ability to reconstruct such image token sequences as high quality and visually diverse images.

Parti has been tested in different complex scenarios and in different variants of photography, comics, oil painting and more, the results were amazing, the artificial intelligence system demonstrated the ability to fulfill the request in various formats and specific styles, with consistent quality by scaling Parti's encoder-decoder up to 20 billion parameters; although not with optimal results, from Google they comment that "While Parti produces high quality results for a wide range of indications, the model, however, has many limitations". So, we assume that it is a model that will be constantly updated until it satisfies the team.

Although for security reasons, Parti is not available for public use, unlike DALL-E Mini, the official project page and the complete research can be accessed.

You can read more about Parti at:

https://parti.research.google/

https://gweb-research-parti.web.app/parti_paper.pdf

26 de Julio, 2022