How we post 30 AI-influencer photos a week without the face changing

You post a great photo on Monday. By Wednesday the jaw has changed a little, the hair is a shade lighter, and the body looks like someone else's. Each photo holds up alone. Side by side, you get three different people.

Viewers catch this faster than most creators expect. People follow a person, not a style, and once they stop recognizing her in the feed, the account loses what made them follow it. If your page is built on one character, consistency is part of the product itself.

What "the same person" means to a viewer

Most creators check the face and stop there. Viewers check three things, usually without knowing it.

The face comes first: eye spacing, nose, jaw, a mole or freckle pattern. It has to survive different angles, distances, expressions and lighting, all the way down to a thumbnail.

Hair is second: color, length, parting, how it falls. Real people change their hair, so a new style is fine when it looks deliberate. What gets noticed is drift that looks like an accident.

The body is the one creators forget until a commenter says she looks different. Height, build and proportions drift quietly between runs, and they are the hardest thing to repair afterward. It gets a full article of its own later.

A fourth layer sits behind all three: the camera look. One glossy studio-style shot in the middle of a feed of phone candids breaks the illusion as badly as a new nose.

LoRA or no LoRA

Without a LoRA, every run gives you a new face. You can describe a person or feed in a reference, but you cannot pin her down. Run it ten times and you meet ten people. Any consistency you get is luck plus rerolling.

That is fine in several situations. A new face every run is the cheap way to test a niche and find out which scenes, hooks and looks people save and share. It also suits formats where nobody expects the same person twice, like a meme hook followed by a reveal, and it needs almost no setup.

A custom character LoRA is a small add-on trained on one specific character. Once it exists, the same person shows up in the gym, at the pool, in a hallway, on a night out, across outfits and light. Generation time per photo stays at about a minute. What changes is how many of those photos have the right face.

Album-style carousels, where five frames of one day have to show one person, are where a missing LoRA shows up first.

Build one when:

• the account is built around one character and followers expect to see her again
• you post often enough that rerolling for a matching face would eat your week
• you already know the niche works, because you tested it without a LoRA
• the character is the asset, as in a "same character, every scene" series

Skip it while you are still guessing what the audience wants, or when you only need the odd image. A fair rule of thumb: if you expect to post the same face every week for more than a month, build it. If not, stay without.

One caveat. A LoRA only learns what its training images show, so one trained mostly on face crops will lock the face and leave hair and body to chance. Covering all three takes deliberate work when you build the character, and that part is in the kit.

How a weekly batch runs

Plan scenes, not photos. A week starts as a list of places and moments: gym mirror, hallway, pool, a night out, a quiet morning. The person stays fixed and the scene changes. We pick scenes from references that already perform on the platform, and we will cover how in a separate piece.

Before the first batch, write down what never changes (face, hair color and length, build) and what is allowed to vary (outfit, makeup, hair up or down). Every later batch gets checked against that page.

Then one session does the whole job. We generate with Krea 2 on a rented RunPod GPU. The pod is ready in about 90 seconds, you run the batch, and you shut it down. Each photo takes about a minute and a half and about $0.04 of GPU, so 30 photos is roughly half an hour of generation and about $1.50.

Review the week as a set. One at a time, almost everything looks fine. Put 30 thumbnails in a grid at phone size and the problems jump out: the face that is slightly narrow, the shot where the body is off, the one that looks like a different camera. Keep a handful of approved photos as your yardstick and compare each batch with them. Whatever breaks the grid gets cut and regenerated.

What survives becomes carousels of three or five images. Thirty photos is ten carousels of three, about a post and a half a day. The rhythm we aim for on TikTok is three posts a day, which at three or five images per carousel means roughly 60 to 100 photos a week. The routine stays the same; the session just runs longer.

What changes when you post at scale

Quality control turns into picking instead of repairing. You choose the best 30 from a batch rather than rescuing photos one at a time, and a rejected photo costs about four cents.

Cost becomes predictable. At about $0.04 a photo, even 100 photos a week is around $4 of GPU. There is no graphics card to buy.

Your attention moves to scenes, hooks and captions, which is where growth comes from once identity is solved. And because followers recognize the character, each post builds on the previous one instead of starting from zero.

Mistakes we see

• Judging consistency one photo at a time. Drift only shows in a set.
• Locking in a face before you know the niche works.
• Changing the scene, outfit, light and framing all at once, so you cannot tell what broke the face.
• Watching the face and ignoring hair and body.
• Mixing looks, such as studio-clean photos in a phone-candid feed.
• Skipping the AI label where the platform asks for one, or posting in bursts that read as spam. Follow the platform's rules on both.

Want this pipeline on your own RunPod, with your own character LoRA? CANDID — link in my profile.