OpenAI is bringing AI image generation directly into its flagship chatbot. Using ChatGPT’s 4o model, you can now natively create AI images in your normal ChatGPT window — no need to use Dall-E. This new feature is available now for free and paying users, with plans to bring it to those on enterprise and education plans next week. You can also try out image generation in Sora, the company’s AI video generator.
OpenAI’s foray into image generation thus far has been focused on Dall-E, a separate program you can use through ChatGPT. You can use Dall-E to create everything from scenes from a sci-fi space opera to stock photography-style shots. Dall-E is our top pick for the best AI image generators, partly because its unique conversational style makes creating and editing images easy. Luckily, that “chat to edit” ability is coming to ChatGPT, too. (Dall-E will still be available to use.)
ChatGPT is available for free, with paid plans offering more features starting at $20 per month. The limits of your current plan will apply to image generation — if you’re on the free plan, you may run into limits using the 4o model for messaging, file uploads and data analysis. The same goes for Sora users. ChatGPT Plus users will get one image per prompt.
Image generation in ChatGPT 4o will focus more on creating work-related images, like infographics and diagrams. OpenAI says it’s improved text rendering to make that happen — something extremely necessary as AI consistently hallucinates and messes up words in images. You can also upload your own images and edit them with AI.
In an example image from OpenAI, the text in this image is remarkably clear.
There are some serious limitations to ChatGPT’s ability. Most importantly, it says that you may not be able to precisely edit specific regions of an image — an essential task as AI models can hallucinate things like eleven-fingered hands. If you upload your own image and make edits to a subject’s face, those edits may be lost from edit to edit. You may also see issues with cropping and struggles with data visualizations and multilingual text. The company says in a blog post that it is working on improving these things and hopes to introduce fixes as early as next week.
Another example of ChatGPT’s image generation skills, featuring a snail-themed pun.
Like Dall-E, images made in ChatGPT don’t have any visible watermarks denoting they are AI-generated. OpenAI said that its images will have C2PA metadata, an industry standard that lets folks know behind-the-scenes that an image is made by AI. In terms of safety, OpenAI says it will follow the same content guidelines as the rest of the 4o model. It said it has “heightened restrictions” around nudity and graphic violence.
For more, check out our full review of ChatGPT and guide to writing the best AI image prompts.
Read the full article here
