All images are generated using independent, separate API calls. See the FAQ at the bottom under “Why is the number of attempts seemingly arbitrary?” and “How are the prompts written?” for more detail, but to quickly summarize:
In addition to giving models multiple attempts to generate an image, we also write several variations of each prompt. This helps prevent models from getting stuck on particular keywords or phrases, which can happen depending on their training data. For example, while “hippity hop” is a relatively common name for the ball-riding toy, it’s also known as a “space hopper.” In some cases, we may even elaborate and provide the model with a dictionary-style definition of more esoteric terms.
This is why providing an “X Attempts” metric is so important. It serves as a rough measure of how “steerable” a given model is - or put another way how much we had to fight with the model in order for it to consistently follow the prompt’s directives.
I had the same experience. Plus you never know what's the best way to use eg. Nano Banana -- it works better in AI Studio versus their regular Gemini chat.
