LoRA Training for Style Consistency
The test: is it still the same character three story branches later, made by someone else? LoRAs per model, each evaluated before production, made the answer yes.
- Client
- Series Entertainment
- My role
- Model training and pipeline design
The test
One good-looking image proves nothing. The real question is whether a character is still unmistakably that character three story branches later, made by someone else, a month on.
What I did
I trained custom LoRAs per model for characters and environments, built so a team could use them rather than a single specialist, and put every one through an evaluation pass before it reached production.
The answer
Yes, across thousands of images and every branch of the story. That is what made the volume usable: a hundred thousand images are worth nothing if someone has to check each one for drift.
