CyberJungle, the Youtube channel of Hamburg-based Senior IT Product Manager Cihan Unur, recently posted a great video on consistent generated characters.

There are lots of great insights in this 20-minute video. Two outstanding takeaways:

First: a prompting guide for Flux.1. At 15:28 he reveals three prompting styles: list, natural language and hybrid.

Second: a guidance guide for Flux.1. At 17:18 he shows Photorealistic and Cinematic images with a wide scope of guidance values. He posits:

“The essence of guidance setting is a compromise or a balance between photo realism and prompt understanding.”

See 18:36 for the Photorealistic results. He prefers a level of two.

See 19:54 for the Cinematic guidance level he prefers: again two.

My take: to me, too often generated images look over-the-top and so ideal, they’re unrealistic. The key seems to be dialing the guidance down to two. Who knew? Now, you do.