Monday, April 1, 2024

AI Image Generators Stumped: Can't Create Simple White or Black Backgrounds, Researchers Discover

Many generative AIs like MidJourney and Dall-E can generate perfectly crafted complex images but it was recently found out that these AI image generators are not good enough for generating simple images. A data scientist, Cody Nash, experienced AI struggling with generating a plain image when he asked the generator to create a White painting. Nash just wanted AI to produce an image with a plain white background but the AI generator couldn't perform such a simple task.

This experience led Cody Nash to write an detailed analysis, explaining whether AI is capable of creating an image with just plain color or not. He asked Dall-E 3 to generate an image with nothing but white color. Dall-E kind of understood the task but couldn't deliver the exact image needed. It added different shapes, images and features to a white background. This was not what Nash needed so he asked Dall-E 3 again to generate an image with nothing but smooth and plain white background. The results were close but not exactly right.

AI Models Fail to Deliver Plain White Backgrounds: Bing Image Creator, Stable Diffusion, Dall-E, and More

Nash told a media outlet that the purpose of his research was to see if AI could generate plain backgrounds. He added that it's good to see that AI is getting creative with its images because even though he was asking for something else, AI was giving him something more creative.

Microsoft Bing Experiment Reveals AI's Struggle with Generating Plain White Backgrounds

After Cody Nosh’s experiment with Dall-E 3, we also ran some experiments asking for a plain white background on Bing Image Creator, a few Stable Diffusion based tools and Dall-E. Not surprisingly, the images generated by these AI models were also not plain white. MidJourney suddenly froze, Microsoft Image Creator took way too much time, while Dall-E gave images which were not related to the prompt. All of them also gave the same results when they were asked for backgrounds of different colors but some are a little better than others. MidJourney started with generating images that were closer to the prompt, but continued adding shapes and features on the top.

We wondering if Google Gemini will do the same, but surprisingly it produce the most accurate result for plan black and white backgrounds.

Google Gemini Surprises us with Accurate Results in Plain Black and White Background Generation

However, it also struggled when we asked about other colors, this prompt for green color "create image a plain background of green color of this code #22B14C and nothing else" produced unwanted results.


This shows that even though AI tools are making our lives easier, there are still some areas where they lack. Not everything can be perfect and that's the deal with AI tools too.

Read next: 65% of Start Ups Are Now Paying for ChatGPT
by Arooj Ahmed via Digital Information World

No comments:

Post a Comment