The development of AI artificial intelligence is advancing rapidly, and it has demonstrated the ability to surpass humans in many aspects. AI not only defeated the world chess king, but also defeated the e-sports champion team (for example, Open AI used its self-developed Bots to let the top players of the “Dota 2” game in a game. The first taste of defeat in an exhibition match). Not only that, AI will also write articles. The GPT-2 and GPT-3 texts launched by the OpenAI Research Laboratory generate pre-trained language models. Because they can write articles comparable to human writing, they have become a powerful tool for writing fake news. Now, OpenAI has launched a new generation of DALL·E 2 system, which can ask AI to generate various pictures for you as long as you pass a description text.
In January last year, Open AI launched DALL·E based on GPT-2/GPT-3 language model and CLIP image recognition system, which can convert user input text into vivid surreal pictures. For example, the user can ask DALL·E to generate a picture of an astronaut riding a horse in outer space through a text description, or a picture of two teddy bears working on new AI research on the moon, so the surreal level is even comparable to the super-realistic Realist painter Salvador Dalí. The word “DALL·E” is a combination of the names of Dali and the Disney movie “WALL-E” (WALL-E).
But the first-generation DALL·E picture pixel is only 256×256, and now the second-generation DALL·E 2 picture quality can reach 1024×1024, so the performance of resolution and low latency is better. Now DALL·E 2 has updated the CLIP system and renamed it unCLIP. The new system supports a process called diffusion, which starts with a pattern of random points and transitions asymptotically into a picture once a more specific focus is captured.
In addition to generating new images, users can also locally change part of the existing images through DALL·E 2, such as adding a duck to the pool or removing an object. The system also incorporates factors such as shadows, reflections, and materials. Consider. Users can also creatively generate additional variant pictures with different styles, contents or angles based on the original pictures.
Just as language models can be used to generate fake news, image generation tools like DALL·E 2 can be abused. In this regard, OpenAI provides some protection mechanisms in place, including that users cannot generate portrait photos based on names, nor can they generate or upload objectionable content. Furthermore, in addition to topics such as hatred, harassment, violence, self-harm, nudity, and illegal activities that are strictly prohibited, it is also prohibited to generate images related to fake news, political situations, medical care, and even diseases.
In the future, Open AI may not directly publicly launch DALL·E 2, but will provide it to third-party apps.