Below we show results of our method together with various models discussed in the paper, showcasing 100 randomly selected samples in a random order from the Waldo & Wenda test set. For models supporting beam search, the top five beams are displayed, and while the images are downscaled for convenience, the results retain the quality of the original images. For externally hosted images, a URL is provided instead of displaying them directly. To explore the results, simply select a model and use the "Previous" and "Next" buttons to navigate through the samples.