Same girl, different scene: How to create an AI influencer​

This is an advanced tutorial. We recommend the beginner guides first

Note: I wrote this a year ago, and the image quality below does not match the current quality, but the same idea still applies.

This is 2024 level of quality (will update the tutorial soon)


So the question I always get is…

Q: I’m able to create the girls I want, but they look like different characters from one scene to another. How do I fix this?

We can definitely help.

All the tools are here.  Try the tactics in this guide, step by step.

Step 1: Lock the character in

Click advanced, and paste this template:


OK, but what is a seed? 

A seed is a made-up reference point. The number is random, but it acts as an anchor, just like GPS coordinates.

The seed number itself doesn’t mean anything per se — it doesn’t refer to a specific person.  It doesn’t mean to look up a specific photo in a database, either.  What seed does is help the AI understand that you’re talking about the same pattern across different prompts, in an abstract way. Seeds are great especially useful when re-running a prompt, and sharing one on the web. Seeds were built for repeating similar results.

For example, “12345” or “3948394” or “6969” are three examples of valid seed numbers. Use whole numbers without zeroes. You can go into the trillions if you’d like, but a bigger number doesn’t make it better or worse, it’s just a random number.

FIGURE 1: How to use Seed

Step 2: Locking the characteristics

Word order matters. A lot.

Describe your girl at the start of your prompt with very specific details: Age, Race, Country

Putting age and some ethnicity or other specific characters at the start of the prompt also makes a big difference. Let’s say I want to create a “20 year old half-Asian woman”. But what kind of Asian?

The more specific we are, the better the chances the AI can repeat that result.  So “20 year old half-Korean red-haired woman” is the better choice. You can also wrap the most important characteristics in nested positive prompts to really drive this home. Now, that’s already a pretty unique idea for a character, let’s see how we did.

So far, our prompt looks like this

 /seed:342342 (A 20-year old half-Korean red haired woman)

Don’t worry about things like size and resolution or quality at first, this just slows the render down. You can boost things at the very end when you have the perfect repeatable prompt. Select Draft to get back 9 images fast. Render your idea and see how you did.

FIGURE 2: Our first draft


Step 3: Reducing inconsistencies

Imagine you’re talking to a police sketch artist. This girl just stole your heart and your wallet. You tell the police that “she looks like a 20 year old Korean college student with red hair”…is that specific enough to catch a crook?

Figure 3: Close but too much variation

Here’s the lineup. They’re all similar, but notice sometimes the hair has bangs, sometimes its long and parted, the nose bridge is different sometimes, and the general shape of her face is sometimes triangular or heart-shaped.  It’s my fault, my prompt sucked. I wasn’t specific enough.  Remember, a seed is just a general pattern, and not a specific person. So we are 50% there.

Go way more specific

Does she have big nose? A pointy chin? Long eye lashes? There’s so many possible chances to arrest the wrong girl just based on that information. We still haven’t really narrowed it down for a human sketch artist, much less for the AI, so inconsistencies are still possible. 

Of course, you can prompt all of these characteristics every time, but the longer the prompt is, the less the AI listens to the words at the end, we still haven’t included things like clothing and what’s happening in the scene. Long detailed prompts are good for creating single images, but not a series of them. So where do we go from here?


Low-rank Adaptation models (LoRAs) are an instant fix to getting what you want. Our system has 1100+ of these in every style, whether it is a likeness of a celebrity or a shortcut to getting someone in certain clothes or in a pose. You don’t have to use celebrities at all, there are hair, clothing, fashion, all kinds of models in here:

Figure 4

Step 4 (Optional) add a resemblance

Unless you’re making your own AI model (we’ll get to this later), the next best thing is blending characteristics from other models, to strongly inherit the same traits every time you want to make an image.

Here’s how to do it:

  • Click Advanced
  • Paste our prompt again, same characteristics, same seed, same base art style  /seed:342342 (A 20-year old half-Korean red haired woman
  • Click Add Concept
  • Click the NSFW tab to reveal all 1100+ models
  • Click on Females
  • Add 1 or 2.  Three tops is recommended. Save space for pose and clothing slots.
  • Use the slider to increase or lower the effect
  • Low weights are recommended to start with. LoRA characteristics are stronger than your prompt. You may need to add more (positive) brackets after a LoRA is applied.
  • Optional:  Change the base art style if you want.  You can combine a recipe with a different base concept.  I went with Realvis13, for example.


<gadot:0.15> – she will look 15% like Gal Gadot

<koreandoll2:0.47> – she will have 47% of the face and body a “Korean Doll”, a premade curvy hot Korean AI girl

Figure 5


Let’s look at the resulting images. Much more consistent already!

As you can see, compared to Figure 2, our results are a lot more consistent already across six images. They’re almost certainly the same person in every image, and we didn’t have to do a lot of prompt engineering for this, either. We didn’t have to prompt anything about parted hair, straight to wavy hair, large breasts, bright red lipstick, and painted eyebrows, those are all characteristics strongly inherent in the LoRas and base model that I selected.

I rendered this using 15 steps, so the realism and detail isn’t as high as it could be.  At this point we’re still developing the basic concept, so we’re not shooting for quality yet.

Time-Saving Tip:

Hit the “Prompt” button under the photo to get the exact prompt that generated the picture.  You can then easily tweak the variables without having to go through the visual menus every time.  Just paste this into the positive box and click the recipe or pick a different recipe if you’re not 100% loving this art style.

Click the Prompt button from your render (or) My Archive screen to continue to Step 5, poses.

Notice that she looks a tad older than my original prompt. This is because the LoRAs I selected influenced her age.

We can stop here and make some creative decisions.  If the base model isn’t doing it for you, try a different one, or make the age a stronger part of the prompt to get what you want.  I can say (((20 year old))) and she will look younger than this image, but still similar. And that is how you can tweak things like hairstyle or lipstick without changing the character.

Common mistake: changing the base model or recipe and expecting the same girl. 

Each AI model was created by a different model maker, so the character is unlikely to appear across different models.  The same prompt will produce an entirely different person if we switch from <realvis30> to <chilloutmix> because the training photos for each model are totally unrelated. The same prompt on a different base model will be a completely different girl.

Step 5: Same Girl, Different Pose

OK, let’s get to the good stuff, how about some fellatio?  

Copy the prompt and add “sucking a man’s dick, oral sex, intercourse” etc.  The more specific you are, the better.

Tip: Be specific when creating sex scenes. If you simply say “sucking a dick” the AI will take that literally and make a dick appear out of thin air. To help you get this right, there are also LoRAs for poses. Go back to the concepts browser, turn on the NSFW toggle, click porn, and add the pose to your prompt.

Here, I turned on NSFW, selected “porn” first and then “oral”.


When you add poses they also have a slider (weight) to control how strong the effect should be.  Starting from 0.7 or lower is recommended. If you get glitches try a different value. Some LoRAs do contain face data, so sliding it to high values may impact your character.

Still pretty consistent results!  Same nose, same hair, and so on.

How about a change of clothes and a different situation next?

Step 6: Different clothes, different situation

Put it all together.

Oh, the places you will go!

Notice how much room we still have in this prompt to describe a much more specific character, but the results are generally consistent. The short hair bangs and high nose bridge never came back, but I aged her a little by mentioning the old men on the cruise. To avoid that, I should have inpainted the old men instead, isolating those words from my girl.

/seed:342342 (A 22-year old half-Korean red haired woman), wide shot, (floral dress), eating a sandwich on the rail side of a luxury cruise out at sea, old men in the background, sunny cloudy sky <gadot:0.42><koreandoll2:0.69>

Remember, we’re still looking at draft quality here. Go back to your prompt and try running it at a higher resolution, using HighDef tools and Facelift.

Even at 512×512 she’s pretty cute:

TLDR Review

Use these tactics to create consistent, repeatable AI characters:

  • Make up a number and add it to your prompt, aka a “Seed” number
  • Use very specific language to describe them at the beginning of the prompt, such as age and nationality
  • Stick to one base model and one aspect ratio when possible
  • Keep a journal of prompts you liked, don’t lose the seed number, etc.
  • Clothing, hairstyles, and poses with LoRAs
  • Reinforce by adding other celebrity names in the positive prompt box
  • Use Inpainting when there is more than 1 character in a scene to separate characters