A clocktower postcard trial cannot be repeated: each run looks different, so staff cannot tell whether a sharper bell came from more steps or from luck. What control should the associate add?
Select an answer to reveal the explanation.
Short Explanation
A fixed seed is like locking the shuffle on a playlist so you can compare equalizers fairly. Without it, sharper bells might just be luck.
Full Explanation
A clocktower postcard trial that cannot be repeated because each run looks different should fix a random seed so the next quality comparisons are reproducible. Civic heritage postcards for public sale and education need fair step or recipe A/Bs; without a seed lock, sharper bells may be luck rather than a better denoising choice, wasting volunteer judgment and municipal creative budget. Reproducibility is the control that makes diffusion quality experiments interpretable for taxpayers.
Disabling all seeds so every run is forced to differ fails by maximizing irreproducibility instead of adding control. Replacing seeds with NCCL ring ordering as the image control fails by importing collective-network ordering into an image generation experiment. Measuring only how fast Kubernetes schedules the pod fails by substituting scheduling latency for visual quality comparison.
Exam caveat: associate diffusion quality comparisons need a fixed seed for reproducibility; fabric ordering and pod schedule speed are not image controls. Operational check: set and document one seed in the trial charter, rerun the prior step/recipe pairs, confirm visual differences stabilize, then proceed with seeded quality sweeps only.