The Computation of Quantitative results

[zihui0930]

Dear Authors,

Thanks for your great work, also with a perfect code implementation. Here I have two questions about the computation of quantitative results, e.g. FID.

1. What is the number of generated or real images used in quantitative results computation? For example, to compute FID, we need a set of generated images and a set of real images. I didn't find the number of them.

2. How to select the generated images or real images? This is the following question: when you take some real images to compute FID, will you select them or simply randomly select some, and what is the source of real images, like train/val/test set? Similar to generated images, I can see that the generated 3D scenes will be rendered for a top-down view, but not clear about the text prompt. Will you use text prompts in the train/val/test set, or randomly generate them?

Thanks,
zihui

[Neal2020GitHub]

Same question.

[2741913295]

Same question

[wmingdao]

Same question