There is no indication that we don’t have enough images already available and training time is the main bottleneck. Also openai clear showed that DALL-E could generate avacado chair even though there is nothing like that in the dataset.