Sampling denoising steps across all stages reportedly scores best on motion and anatomy

A commenter describes a 100-prompt test using the same reward model: early-only denoising-step training scored better on motion than late-only training, while late-only scored better on anatomy. Sampling across all stages scored best on both. They say the results support the paper's overall system but don’t isolate the generator backbone’s contribution, and call for a comparison using a different full-video reward model with the same ReFL training setup.