I’d like to ask whether the evaluation for vbench-long uses clip_length_mix.yaml for all dimensions?

When using clip_length_mix.yaml, it seems that
imaging_quality: 2.0
aesthetic_quality: 2.0
are actually not being used!

In the code:

length_values = [clip_lengths[dim] for dim in unique_dimensions if dim in clip_lengths]

Since in VBench_full_info.json, these two dimensions are grouped together with overall_consistency, running

python vbench2_beta_long/eval_long.py --videos_path xx --dimension aesthetic_quality

will always result in
length_values = [2.0, 10.0, 2.0],

so

max_length = max(length_values) if length_values else None

will always be 10.

Therefore, for the imaging_quality and aesthetic_quality splits, the evaluation will always use the 10.0 length setting.

I’m wondering if this is the intended behavior, or should it be adjusted manually?