Likeness consistency across 12 shots without a trained LoRA — reference-image workflow

by @chenwagner_creates485 in c/portrait-studio · · Character

Needed the same character across a dozen portraits (different lighting, expressions) for a project, didn't want to train a full LoRA. Here's what held likeness best:

Lock ONE hero shot you love. Neutral expression, even light. This is the bible.
Every subsequent shot: reference-image the hero at medium strength, and only change ONE variable (lighting OR expression OR angle), never two at once. Two variables and the face drifts.
For profile/3-4 angles, do them as an img2img from a front shot, not a fresh gen. Fresh gens of 'the same person from the side' are a coin flip.
Final unify pass: once you have all 12, pick the 3 that drifted most and regenerate referencing TWO of the good ones, not just the hero. Averages them back to the group.

Not as locked as a LoRA but 90% there for a tenth of the effort. Soul's reference adherence made this way easier than it would've been a year ago.

541 points · 5 comments

Comments

  • @elenajohnson_films1849 · · 71 points

    The 'regenerate the drifters referencing two good shots to average back' trick is genuinely smart, stealing it. Treating the set as a population you pull outliers back toward.

  • @nikolaipatellab · · 99 points

    'Change ONE variable at a time' is THE rule for likeness and nobody follows it because they want the dramatic side-lit laughing 3/4 shot immediately and the face becomes someone's cousin.

  • @kenjimartinmedia · · 62 points

    Guilty. I always try to change lighting AND expression AND angle in one gen and then act surprised when it's a new person. Discipline is the whole skill.

  • @dianasilvaworks5338 · · 43 points

    Reference workflow vs LoRA: for under ~15 shots, reference wins on effort. Past that, train the LoRA, the consistency ceiling is just higher. Good rule of thumb buried in here.

  • @skylergarcia_art · · 29 points

    Agreed. The crossover is around 15-20 images for me too. Below that LoRA training overhead isn't worth it.