OpenAI has announced the third version of its artificial intelligence-powered image creation platform DALL-E, which now allows users to utilize ChatGPT to generate instructions and includes additional safety features.
DALL-E converts text instructions into images. However, even DALL-E 2 made many mistakes and often ignored specific phrasings. The latest version, according to OpenAI researchers, understands the context much better.
A new feature of DALL-E 3 is integration with ChatGPT.
### What’s New in DALL-E 3
Through ChatGPT, there is no need to invent a detailed instruction; you can simply ask ChatGPT to come up with an instruction, and the chatbot will write a paragraph (DALL-E works better with longer sentences) for DALL-E 3 to continue. Other users can still use their own instructions if they have specific ideas for DALL-E.
In a demonstration to The Verge, Aditiya Ramach, lead researcher and head of the DALL-E team, asked ChatGPT to help him invent a logo for a mountain restaurant. ChatGPT then wrote a longer instruction, and DALL-E generated four options. The chatbot, said OpenAI, allows more people to create AI art since they don’t need to be very good at coming up with instructions.
[Watch the video here](https://youtu.be/sqQrN0iZBs0)
### Differences Between Various Versions of DALL-E
DALL-E, first released in January 2021, arrived before other platforms of AI art text-to-image generators like Stability AI and Midjourney. Until DALL-E 2 was released in 2022, OpenAI had a waiting list to control who could use the platform after criticism that DALL-E could create explicit photorealistic images and showed bias when creating images. The company removed the waiting list last September and opened DALL-E 2 to the public.
This new version of DALL-E will be released first to ChatGPT Plus and ChatGPT Enterprise users in October, followed by research labs and its API service in the fall. OpenAI plans to amaze with the release of DALL-E 3 but has not committed to a specific date for the release of a free public version.
### New Security Measures
OpenAI claims to have focused heavily on creating strong safety measures in DALL-E 3 to prevent the creation of inappropriate or potentially violent images. OpenAI said it worked with external red teams – a group intentionally trying to break the system to test its safety – and relied on input classifiers, a way to teach language models to ignore certain words. DALL-E 3 also cannot reproduce images of public figures – provided the instruction explicitly mentions the name.
Sandhini Agarwal, a policy researcher at the company, said she has “high confidence” in the safety measures but clarified that the model is constantly improving and not perfect. OpenAI representatives said DALL-E 3 was trained to refuse to create images in the style of living artists. Unlike DALL-E 2, which can mimic the style of specific artists when requested.
To avoid legal claims, OpenAI will also allow artists to revoke their art in future versions of text-to-image models. Creators can submit a takedown request for an image they own on their website. A future version of DALL-E could then block results that look similar to the artist’s image and style. Artists sued DALL-E competitors Stability AI and Midjourney, along with the art site DeviantArt, for allegedly using their copyrighted works to train their text-to-image models.
