The DALL-E neural network is capable of creating images from text descriptions. Now the developers have presented an updated DALL-E 2, which can edit ready-made pictures.
DALL-E is based on the GPT-3 architecture and can generate images by processing English text queries. The company first introduced it in January 2021, where it said that AI understands complex abstract combinations and even learned the historical and geographical context.
Now OpenAI is sharing the new features of DALL-E 2: drawings have become more detailed and realistic, and now you can change fragments on finished images or create variations of the original image from different angles or styles. For example, you can edit a photo and replace a dog with a cat on it, or offer a couple of options for a canvas by a famous artist.
As described in the blog, thanks to deep learning, the neural network not only understands individual objects (for example, “koala” or “motorcycle”), but also learns from the relationships between them (and when querying “koala rides a motorcycle”, the neural network knows how to connect it with other objects or actions, and will suggest drawing).
At the same time, possible errors of the neural network are also noted. For example, if during training the picture was named incorrectly (a car was indicated instead of an airplane), then when requesting “aircraft”, DALL-E 2 can draw exactly the car.
So far, closed testing of DALL-E 2 is being carried out. The company offers to join the waiting list for access to the neural network.