AI influencers: a campaign at the snap of your fingers
Character system for influencer marketing
AI Influencers is a system for influencer marketing inside ZDES Studio. You create a character once, shaped around the brand’s audience, and from there they appear in any scene — at a café, on a run, at work, out with friends — and stay recognizable in every shot, clip and story.
- Product Owner
- Character pipeline
- Prompt architecture
- Art direction
- Launch
01
A whole system, not an image generator
A typical influencer campaign means scouting creators, rounds of approvals, shoots and a lot of waiting. A typical image generator gives you a beautiful shot of a slightly different person every time. I designed something in between: a multi-stage pipeline where every step has one job.
The avatar is created once, in four steps. You upload photos of the character, up to fourteen, and pick the main one — it becomes the identity anchor. The system analyzes it, builds an identity profile and fills in the settings on its own: gender, age, build, hair, eyes, face shape. All that’s left is to adjust them and add distinctive features — a mole, freckles, a scar on the brow. You can skip the photos too and describe the character in words. Create avatar produces a base portrait, and the avatar goes into the library.
From there you work with the avatar in tabs. The main one is Shoot: optional background, style and product references, a few words on the direction and the number of shots. The system plans a series in one location and one visual style, each shot with its own role: lifestyle, portrait, with the product, detail, mood. The other tabs handle specific jobs: Editing changes the look, outfit, pose or background, Product puts the avatar with an item, Talking Video adds a voice of its own and talking video, Socials lays shots out for each platform, and Together puts up to three avatars in one frame.
02
How it works
Two stages. First the avatar is created once, from photos or from a description. Then you work with it in tabs: Shoot produces a series of consistent shots, the other tabs handle edits and formats. Step names follow the product’s interface.
flowchart TB
subgraph create ["Creating the avatar"]
direction LR
photo("Upload photos<br/>up to 14") --> main("Main photo<br/>the identity anchor") --> analyze("Photo analysis<br/>face · hair · eyes") --> identity("Identity<br/>settings from the photo") --> base("Create avatar<br/>base portrait")
manual("Describe from scratch<br/>no photos") --> base
end
subgraph use ["Working with the avatar"]
direction LR
tabs("Other tabs<br/>Editing · Product<br/>Talking Video · Socials · Together") --> out("Edits · product shots<br/>talking video<br/>11 platform formats")
shoot("Shoot<br/>references · direction<br/>3–10 shots") --> pipeline("Identity · style<br/>creative direction<br/>prompts · face lock") --> series("A series of shots<br/>Lifestyle · Portrait<br/>With product · Detail · Mood")
end
create -- "the avatar lands in the Avatars library" --> use
class base,shoot key
class series result
classDef key stroke:#FF3600,stroke-width:2px
classDef result fill:#334cdb,stroke:#334cdb,color:#ffffff03
Prompt architecture
Shoot is a chain of five steps, P0 to P4. I trusted the model only with creative decisions: the concepts and the technical prompts. Everything that touches identity is assembled by code. One image defines the face, the avatar’s main photo, and no prompt can paraphrase it.
- 01code template, no LLM
P0 · Identity descriptor
The photo analysis made when the avatar was created (temperature 0.2, “if it isn’t visible, don’t guess”) and the confirmed settings are combined into an IDENTITY DESCRIPTOR block. There is no second vision call, so the identity text stays the same across campaigns.
- 02LLM with image · 0.5 · only with a style reference
P1 · Visual DNA
When a style reference is uploaded, the model breaks it down along six axes: palette, lighting logic, texture, framing and camera, cultural references, emotion. Without a reference the step is skipped.
- 03LLM · 0.6
P2 · Creative director
Writes N concepts, one frame each, as frames from a single shoot in one location and one light. Pose and expression variety is mandatory: without it a locked face freezes into one pose. The product is spread across the series, leading in some frames and in the background in others.
- 04LLM with images · 0.5
P3 · A.O.C. architect
Turns the concepts into technical Anchor / Optics / Chemistry prompts. The order is strict: first the person from Image 1, then the environment, the style and the product; in a conflict the earlier step wins. The model must ignore any people in the other references.
- 05code template, no LLM
P4 · IDENTITY LOCK
The code assembles the final prompt: the reference map, the IDENTITY LOCK, a PRODUCT LOCK when there is a product, and only below them the A.O.C. prompt. The numbers in the map match the order in which the images go to the renderer. The model never writes these blocks and can’t retell them in its own words.
- 06set in the P3 prompt
Shot roles
The first five shots follow fixed roles: lifestyle, portrait, with product, detail, mood. With more shots the roles cycle again from new angles, so a series always reads as a feed rather than five identical portraits.
Prompt fragments
Verbatim, exactly as the model sees them.
IDENTITY LOCK: ONLY Image 1 defines the person's identity. If other reference images contain people, COMPLETELY IGNORE those people — use them ONLY for environment/style/product, NEVER for face or body. - Reproduce the person from Image 1 with 100% fidelity - Preserve exact facial structure, proportions, skin tone, and all distinguishing features - No smoothing, idealizing, reshaping, or blending with ANY other face from ANY other image
1. Resolve the person's identity, pose, expression, and body language from Image 1 and the IDENTITY DESCRIPTOR. 2. Resolve environment, composition, and lighting behavior from the background reference (if provided) or CREATIVE DIRECT (if not). 3. Apply aesthetic style, visual mood, and image character from the style reference (if provided) or CREATIVE DIRECT (if not). 4. Integrate the product (if provided) after environment and person are resolved. If conflicts occur, earlier steps override later ones. Identity preservation always takes highest priority.
- All concepts must take place in one single, continuous environment. The location, light, and atmosphere remain consistent across every shot. - Pose and expression variety is essential. The person should feel alive — not posing for each shot identically.
More in the articles
04
Five characters
Anya, Sonya, Lera, Dasha and Vika. Each has her own look and her own feed: the clips play one after another, the way they’d run on her socials.
05
The same person in every shot
What makes an AI influencer work is recognition. If the face drifts from clip to clip, the audience stops buying the character. So in the system the character’s identity outranks everything else: the base portrait and identity profile go into every generation, and location, style or product references can never override the face. A series is shot in one location and one visual style, and each clip is animated from an already approved still, so the character stays the same in motion.
The clips look the way real creators shoot: handheld, on a phone, with a little shake and natural pauses, none of the ad gloss and no slow-mo.
06
Campaigns at the snap of your fingers
A single run produces a series of three to ten consistent shots of one character. That’s the raw material for a whole campaign: posts, stories, YouTube, Reels and Shorts covers, a five-slide carousel, a 3×3 grid for Instagram and TikTok — eleven platform formats in all. Up to three characters can share a frame, a character can hold the brand’s product, and images come out at up to 4K. Need a new scene or another series? The character is ready — just send the brief.
A producer runs the campaign: scripts, picking the best takes, making sure the character stays themselves. The tool makes repeatable what used to depend on a lucky generation.