Aesthetic Score Computation

Hi Trellis Authors!

My collaborators and I have thus far truly enjoyed reading through your work and playing around with your repository. We had a couple of questions in regards to reproducing the Aesthetic Score Computation:

Rendering

• In the paper, it's mentioned that for the aesthetic score computation you render 4 uniform views. We are wondering how you render these assets? Concretely:

1. Do you use the same rendering script as in dataset_toolkits?
2. What is meant with 4 uniformly rendered views. Do you render 150 first and then select 4 at random, or are there any specific elevations/angles/azimuths that these assets are rendered at?
3. For the aesthetic score computation, do you perform any further processing (i.e. background color/resizing for instance)

Improved Aesthetics Repository
We also had a couple of questions about the linked improved aesthetics repository here.

• Which checkpoint do you use? And what is the model configuration?
• For some reason our scores are always below 5.5, which is quite strange as the same assets in the metadata.csv are listed as 6.6 for instance.I have displayed the distribution of aesthetic scores of the 150 rendered views for the 000045aad61c956b45fc468b2b2ec954636e5f647f1c19 asset below (which has an aesthetic score of 6.55 in the metadata.csv supplied by you)

Start by reviewing the dataset_toolkits rendering approach, the metadata.csv entry for the referenced asset, and the linked improved-aesthetic-predictor repository. Determine the four-view camera choices, rendering post-processing, checkpoint and model configuration, and explain the score discrepancy. Done means documenting reproducible settings that account for the reported values.