OpenAI has announced a major update to its text-to-image AI generator DALL-E. The latest version – DALL-E 3 – powered by the AI company’s brainchild, ChatGPT, will allow users to get refined images with better details as compared to previous versions. In addition, DALL-E 3 has added safety measures to help combat exploitation and misuse, especially in areas such as generation of harmful content.
With advancements in the generative AI space, users continue to easily create images through text prompts. However, most models struggle with getting finer details in the image, especially if a user wants a certain text/word in the image. Getting realistic images with the right details might require prompt engineering techniques to best structure text prompts fed to a Large Language Model (LLM) to get the desired output.
Now, with its integrations with ChatGPT, DALL-E 3 seeks to address this by removing the tedious process of having to refine prompts to generate desired images.
The new update will allow users to just give ChatGPT the task of coming up with the prompts. Since DALL-E works best with longer prompts, once prompted with an idea, ChatGPT will return a detailed paragraph for DALL-E 3 to follow. If unsatisfied with the generated output, users can still ask the AI chatbot to tweak the image with just a few words.
Users can still generate desired images with their preferred prompts.
Addressing challenges with generative AI models
OpenAI has also announced it’s adding safety measures to DALL-E 3 to help prevent misuse and spread of misinformation. Earlier this year, fake images of former U.S. president Trump arrest went viral on social media. The images generated by MidJourney, another text-to-image generative AI platform, despite being easy to spot errors showed how easy it would be to spread misinformation online.
Given how fast the field is evolving, it is important to have mitigations in place to combat exploitations of generative AI models. OpenAI claims it has improved safety performance in “risk areas” including harmful biases and generation of public figures. It further adds that DALL-E 3 will not generate images of public figures when prompted by name.
In addition, the company is working on a tool that will help identify images generated by DALL-E 3. The new classifier which is still in the alpha testing phase will help prevent the spread of misinformation and propaganda from images generated by malicious actors. OpenAI had earlier this year released a text classifier aimed at identifying text content generated by AI. However, the text-detection tool has since been shut down due to lower accuracy.
Generative AI models require large amounts of data for training. This may result in major challenges including issues of copyright infringement with companies training their models on copyrighted content.
To address copyright issues and avoid potential lawsuits, OpenAI claims DALL-E 3 will decline to generate images in the style of a living artist. Creators can also opt their images out of the training dataset by submitting a removal request.
DALL·E 3 will be available to ChatGPT Plus and Enterprise customers in early October.
