Consistency Test: I ran the same character through 12 shots with 4 different identity methods. Clear winner, with receipts.

by @lunadavis_vfx in c/character-animation · · Consistency Test

Spent a week benchmarking character consistency methods on ONE designed character across 12 shots (different angles, lighting, expressions). Scored each on face-match, outfit-match, and how much hand-fixing each shot needed. Here's the ranking.

Trained character LoRA (~30 images) — winner by a mile. Most robust to angle and lighting. Outfit held best. Downside: 2 hours to build the dataset + train, and it bakes in the character so changing their jacket means a new LoRA.

InstantID / PuLID (single reference) — great for FACE, bad for everything else. Face match was shockingly good from one image. But it only locks the face — the hair, outfit, and body drift freely. Good for a face-lock LAYER on top of something else, not a full solution.

IPAdapter (reference image) — decent all-rounder, soft. Holds a general 'vibe' of the character but softens specifics. The face is RECOGNIZABLE but not IDENTICAL. Cousin drift central.

'Just describe them really specifically in the prompt' — don't. 'Woman, 30s, red bob, green eyes, freckles' gives you a different woman every single time who happens to match the description. Prompt-only identity is a myth.

The actual answer: LoRA for the body/outfit/style + InstantID on top to nail the face. Belt and suspenders. Full comparison grid in comments.

1,184 points · 9 comments

Comments

  • @mayamartinez_digital · · 96 points

    the InstantID 'only locks the face, everything else drifts' finding matches my experience exactly. the face is uncanny-good from one ref, then the hairstyle changes shot to shot and ruins it. face-lock is a LAYER, not a solution. thank you for naming that clearly.

  • @elenacosta_studio · · 58 points

    the hair is secretly as important as the face for recognition and nothing locks hair well yet. a perfect face under drifting hair still reads as a different person. hair is the next frontier.

  • @marcusgarcia_vfx · · 98 points

    LoRA + InstantID stacked is the meta right now and you just proved it with actual receipts instead of vibes. the LoRA holds the WHOLE character (outfit, proportions, style) and InstantID enforces the face on top. neither alone is enough. this should be pinned.

  • @omarvermalab · · 99 points

    the 'prompt-only identity is a myth' finding needs to be screamed from rooftops. a description is a CATEGORY, not an identity. 'red-haired woman' is a bucket containing infinite different women. the model samples a new one every seed. people waste weeks before they accept this.

  • @harperthompson_digital · · 34 points

    how many images for the LoRA before returns diminish? you said ~30, is 15 viable for a quick character or does it fall apart?

  • @rileyrodrigues · · 61 points

    15 works if they're DIVERSE (angles, expressions, lighting all varied). 30 mediocre same-y images is worse than 15 varied ones. dataset diversity beats dataset size. the LoRA learns what VARIES as 'allowed to vary' and what's constant as 'the identity'. so vary everything except the identity.