I've seen a lot of requests and questions about creating consistent characters. I used this exact process in creating Blair.
Since XL doesn't currently have very many embeddings for people/celebrities/nobodies, this is using 1.5. I've selected SG_161222's Realistic Vision V6.0 B1 since it's the best photorealistic 1.5 I've seen currently. I've select two of the highest rated nobody embeddings to use in this: Zovya's Cherry and Erin
Using the prompt "Cannon EOS 5D Mark III, portrait photo of cherrynobodysd15, highly detailed skin, pores, peach fuzz, natural lighting, natural shadows" with "rmadanegative402_sd15-neg" as the negative prompt, running at 60 steps at 7 CFG, here is an x-grid from A1111.
Prompting [cherrynobodysd15:ErinNobodySD15] alone has each name/embedding applied at alternating steps evenly. So cherrynobodysd15 1st step, ErinNobodySD15 2nd step, cherrynobodysd15 3rd step, ErinNobodySD15 4th step, etc.
This is an example of that method:
Prompting [cherrynobodysd15:ErinNobodySD15:30] like in the initial example image has each name/embedding applied at different points. The tells it to use cherrynobodysd15 for the first 29 steps and then start using ErinNobodySD15 at step 30.
Here is another example, but with swapping the order:
Prompting [cherrynobodysd15:ErinNobodySD15:10] ended up creating the foundation using cherrynobodysd15 and filled in the rest with ErinNobodySD15 because it started using ErinNobodySD15 at step 10.
Flipping it around to [ErinNobodySD15:cherrynobodysd15:10] does the opposite, creating the foundation with ErinNobodySD15 and filling in the details with cherrynobodysd15.
And now we can see that it offers consistency!
I hope this helps!
It's important to note that this functionality does not work with extra network like LoRAs.