Need Help: my character looks perfect from the front but becomes a stranger the moment she turns her head. Profile = different woman.

by @emmapatellab in c/character-animation · · Need Help

Fighting cousin drift and losing. My OC is rock-solid from the front — trained a LoRA, she's consistent shot to shot as long as she's facing roughly camera. But the SECOND she turns to profile or looks away, she becomes a completely different person. Not melted, just... someone else who vaguely resembles her.

What I've got:
LoRA trained on ~25 images
InstantID face-lock from a front reference

I think the problem is my training images are ALL front-ish? Because she's great at the angles I trained and a stranger at the angles I didn't. But I don't HAVE profile references of her that are consistent, because she drifts at profile, which is the problem I'm trying to solve. Chicken and egg. How do I break the loop? Beginner-ish, be gentle.

96 points · 8 comments

Comments

  • @masonsantos_art · · 100 points

    You diagnosed it yourself: the LoRA only knows the angles it was trained on, and InstantID's front reference can't enforce a profile (it has no profile data of her). The fix for the chicken-and-egg is to BOOTSTRAP the turnaround manually. Here's how:

  • @marcushernandezpro · · 100 points

    Take your best front image. 2) Generate a 3/4 view, and instead of accepting the drift, HAND-FIX it in an image editor until it's unmistakably her at 3/4 (paint her actual nose shape, her actual eye spacing, etc). 3) Now use THAT corrected 3/4 to generate and fix the profile. 4) Add all these hand-corrected angles to your LoRA dataset and retrain. You're manually building the consistent turnaround that the model couldn't, then teaching the model from it.

  • @rileysharma_films · · 67 points

    the all-front-facing dataset is THE most common beginner consistency mistake. the LoRA generalizes terribly to unseen angles. you need profile, 3/4, looking-up, looking-down in the training set or those angles are a coin flip every time. diversity of ANGLE specifically, not just lighting.

  • @zarapark_art · · 41 points

    +1, I now deliberately include a few back-of-head and extreme angle shots in every character dataset even though they feel useless, precisely so those angles don't fall apart in the one scene where I need them.

  • @williamwatanabemedia539 · · 59 points

    cousin drift is basically absent across these shots and that alone is state-of-the-art for a solo creator. how many cleanup passes?