Measuring drift across a series: a quick, honest check

"Consistent" is easy to claim and hard to see on a phone screen. Here is a quick way to check a series before you post it. 1. The contact sheet Put every image of the character in a grid at the same size. Look at the grid, not single images. Drift that is invisible one at a time is obvious in a grid. 2. Crop to the face Crop each face to the same framing (eyes on the same line) and place them in a row. Compare eye spacing, nose length, jaw shape, hairline. These are what identity rests on. 3. The feature checklist For each image, tick: hair colour, hair parting side, eye colour, signature item present and on the correct side, apparent age. Count the ticks. A series where every image scores full marks on the list but the faces differ still has drift; the list catches the easy cases only. 4. Flip test Mirror one image horizontally and compare it with another. Asymmetries (a parting, a mole, an earring) that swap sides show the model is not tracking the feature. What to do with drift - Face drifts: raise reference weight or switch to a face-specific adapter; or train a LoRA. - Small features drift: make them bigger, or fix them by inpainting. - Age drifts: put an explicit age in the record line and remove words that imply age ("elegant", "cute"). Post the grid On this board, post the grid and the checklist with your method. "It drifted at image 6 when I changed the lighting" is the kind of finding that helps the next person.