GoKawiilIt used to be easy enough to distinguish between human-made and AI-generated imagery — just two years ago, you couldn’t use image models to create a menu for a Mexican restaurant without inventing new culinary delights like “enchuita,” “churiros,” “burrto,” and “margartas.”
Now, when I ask the brand new ChatGPT Images 2.0 model for a menu of Mexican food, it creates something that could immediately be used in a restaurant without customers noticing that something’s off. (However, ceviche priced at $13.50 might make me question the quality of the fish).
Image Credits:ChatGPT Images 2.0
For comparison, here’s the result I got from DALL-E 3 two years ago. (At the time, ChatGPT did not generate images):
Image Credits:Microsoft Designer (DALL-E 3)
AI image generators have historically struggled to spell because they generally used diffusion models, which work by reconstructing images from noise.
“The diffusion models […] are reconstructing a given input,” Asmelash Teka Hadgu, founder and CEO of Lesan AI, told TechCrunch in 2024. “We can assume writings on an image are a very, very tiny part, so the image generator learns the patterns that cover more of these pixels.”
Researchers have since explored other mechanisms for image generation, like autoregressive models, which make predictions about what an image should look like and function more like an LLM.
Unfortunately, OpenAI declined to answer a question in a press briefing this week about what kind of model is powering ChatGPT Images 2.0.
Techcrunch event Meet your next investor or portfolio startup at Disrupt
... continue reading