The identity edit workflow takes a scene and a photo of a person and returns the scene with that person in it. No dataset, no training, about 25 seconds per picture on a 5090. It is what powers the free preview on this site and the Bulk Reference Swap workflow. So why train LoRAs at all? Because the two methods fail in opposite places, and once you know where, the choice per job is obvious.
Academy / Characters that stay themselves / Lesson 3
Pro lesson · 6 min
Reference photo vs LoRA: when to use which
Krea 2 can put a person into a picture from one photo, with no training, in under thirty seconds. Or you can spend ninety minutes training a LoRA. Both are right, for different jobs.
The rest of this lesson is for Pro members
Pro is $29 for 3 months and includes every lesson, every premium workflow and the whole character roster.