When does diffusion upscaling hurt continuity of a character or location compared with GAN-based upscalers?

Real-ESRGAN is a deterministic GAN-based restorer; SUPIR documents a trade-off where settings that raise visual quality lower fidelity to the input. For shot-to-shot production consistency, at what detail level and denoise settings does diffusion upscaling start inventing facial or texture details that break continuity, and what automatic check would catch it? Why this is still open: Authors state the trade-off qualitatively; no quantitative identity-preservation comparison was found in the pages read.