By skyforbes 
I was testing the accuracy of the image model by giving it some basic illustrations to generate with simple but very specific qualities. The outputs were all generally on the right track in the design, although completely wrong in rendering specific properties. But that is besides the point right now.
I was giving it instructions one after another, like "A rectangular array of black dots arranged in 3 rows and 4 columns, creating a 3×4 grid pattern." (This actually generated a 3×3 black dot pattern)
The interesting thing came in the next generation when I gave the prompt about drawing a 3" line above a ruler, the output of which contained a noisy negative of the previous image.
This means that previous outputs are being used for input to the next generation, presumably to assist the model in making changes to the image instead of generating a brand new one. This makes sense if you have ever tried to make tweaks to a large image and the outputs following the first generation are much faster.
However, this approach may be a detriment in certain cases, if the images and prompts are completely unrelated, as can be seen here.