Most LoRA trainings for person specific generation hit quality issues that make results look artificial. The breakthrough came from rethinking how lighting and expressions are handled during training and adjusting generation parameters.

Here are some examples:

- Corporate photo in office: https://imgur.com/a/kUQtMZr - TedX speaker: https://imgur.com/a/KOC6eCy - Rockstar performer: https://imgur.com/a/B2YaUHj - Sunset portrait: https://imgur.com/a/d4rEANW

Using Flux as base model. Has aynone pushed realism to this level in image generation?