Running an AI influencer that looks the same in every post

Start with a character bible

Before generating any video, write down who the influencer is: name, age range, niche, where they live, how they talk, what they never do. Then fix the visual identity: face, hair color and length, skin tone, height relative to objects, two or three signature outfits, one or two recurring accessories.

Every decision you leave open will be made differently by the model in each video. The bible is also what you show a brand partner when they ask who they are working with.

Build the reference sets

Design the face once with an image model, then generate or edit a set of views of that same face: front, three-quarter left and right, profile, and a close-up. Pick the set where the person is unmistakably the same in every image; one off image will pull every video toward it.

Treat each signature outfit as its own version of the character, with its own full set, rather than describing the outfit in prompts. A typical account runs on three versions: everyday, going out, and a niche-specific look such as gym wear or a work outfit.

Keep a fixed home location too. A recurring apartment, café or studio, kept as a photo you reuse, makes the account feel like one person's life rather than a stock-footage collage.

Freeze the sets. Do not swap images as you go. If you improve the face in month three, the account will visibly change in month three.

A weekly production routine

Plan content in batches. A workable rhythm for a small team: on Monday, script five to seven posts for the week; on Tuesday, generate two to three takes per shot; on Wednesday, edit and add captions in your editor; on Thursday, schedule.

Mix formats so the account does not depend on long talking clips. A typical week might include two short talking-to-camera clips of 5 to 10 seconds, two lifestyle clips with no speech, such as walking, cooking or training, one product-focused post if there is a partner, and stills pulled from the best frames for carousels and stories.

Before publishing, put the new clips next to two older posts. If you hesitate over whether it is the same person, regenerate.

Labels, disclosure and platform rules

Say it is AI. State in the bio that the account is a virtual or AI-generated creator. As of October 2026, TikTok, YouTube and Meta each ask creators to label realistic AI-generated or synthetic content and offer an AI label or disclosure setting when you post, according to their creator help pages. Use it on every video, not only the most realistic ones.

In the EU, the transparency rules in Article 50 of the AI Act have applied since August 2, 2026. Anyone publishing a deepfake, meaning realistic synthetic content that would falsely appear authentic, must disclose that it is artificially generated or manipulated.

Whether a fully fictional persona counts as a deepfake is debated. Disclosing anyway costs nothing.

Sponsored posts need sponsorship disclosure as usual. According to the FTC's Endorsement Guides, revised in 2023, virtual influencers count as endorsers, and the advertiser as well as the endorser can be liable. A virtual influencer cannot have personal experience of a product, so avoid scripts like "I've been using this for weeks"; present products, do not testify about them.

Do not base the face on a real person without their written consent, and avoid looking deliberately like a celebrity. That creates likeness and impersonation problems no label fixes.

Common ways accounts break

Gradual drift: each post is close to the last, but post 40 no longer matches post 1. Always compare against the original references, not last week's video.

Age and body changes between scenes, especially in full-body shots. Include one full-body image in each outfit set.

Hands and held objects. Phones, coffee cups and products are where realism slips most often. Keep held objects simple and the grip visible.

Over-polish. Perfect skin and perfect light every day looks synthetic even when labeled. Vary light and framing the way a phone camera would.

Where RefVid fits

In RefVid, each version of your influencer is its own subject, built from 1 to 5 photos, and the subject's page lists every video made from it, so drift against the originals is easy to spot. A photo of the home location can be added to a video as a scene photo, which sets the place and look.

Clips run 3 to 15 seconds in 9:16, 16:9 or 1:1. Sound is optional: with it on, a line written in quotes in the prompt is spoken in the clip. RefVid has no voice cloning, so the voice is not fixed from one clip to the next.

RefVid runs on Kling 3.0 Omni today and plans to switch to Kling 4 Omni Reference when that model's API opens; your saved photos stay as they are. RefVid is not affiliated with any model provider.