How to Create an AI Influencer: Face, Voice, Costs and Rules

Most AI influencer accounts die in week three. The first ten posts look great because each one was generated from scratch by someone with time on their hands. Then the face drifts, the jawline changes between a carousel and a reel, the voice sounds like a different person in every clip, and the account reads as a collection of stock renders instead of a character.

Creating an AI influencer is mostly an operations problem. What keeps the account alive is a written character spec, a locked identity, one voice, a pipeline that produces a week of posts without manual clicking, a known cost per post, and disclosure that holds up with YouTube, TikTok, Meta, the FTC and, since August 2026, the EU. This guide walks through each of those in order.

Step 1: write the character bible before you generate anything

Every vendor workflow for this starts in the same place. Kapwing's 2026 guide opens with choosing a niche and gathering references, and Higgsfield's version opens with building the character. Both skip the part that matters in month two: a written spec pasted into every prompt, so the character is defined by text you control.

Keep it in a plain file in your repo:

Niche and promise. One sentence: who follows this account and what they get from it. "Budget travel in Southeast Asia for people with one week off" beats "travel".
Fixed physical traits. Age range, face shape, eye colour, hair colour and length, skin tone, one distinguishing mark (a scar, freckles, a specific earring). These never change.
Variable traits. What is allowed to change between posts: outfits, hairstyle within a range, makeup, locations.
Visual style. Camera and light language, for example "handheld phone photo, natural daylight, slight grain". Pick one and hold it.
Voice. Accent, pace, register, and the exact voice ID once you choose one (step 3).
Personality rules. Three things the character always does and three things it never says. It keeps agent-written captions consistent.
Disclosure line. The exact wording that goes in the bio and in sponsored posts (step 6).

The fixed traits block becomes the identity paragraph that starts every image and video prompt. The variable traits become the part you change per post.

Step 2: lock the face across images and video

A written description alone will not hold a face steady across hundreds of generations. You need a reference set, and the approaches differ by tool. Higgsfield trains a dedicated identity it calls Soul ID on 20 or more photos of the character, in 3 to 5 minutes. Kapwing's character builder takes up to 10 reference photos plus a description and a voice. The model-agnostic version of the same idea works with any image model that accepts reference images.

Build the reference set

Generate 30 to 40 candidate portraits from the identity paragraph alone, with a neutral background and even light.
Pick the single best face. This is your anchor image. Everything else is judged against it.
Using the anchor as reference, generate a character sheet: front, three-quarter left, three-quarter right, profile, full body, smiling, neutral, and one in the target visual style.
Reject any image where the fixed traits drift.
Store the approved set with stable names and never regenerate it. New content references these files, never the latest post.

Carry the face into video

For video, the most reliable path is to generate the still first, approve it, and then animate it with an image-to-video model where the model supports a first frame. Higgsfield's workflow does exactly this, with image-to-video clips as the step after locking identity. Keep clips short, because identity tends to drift the longer a model runs without a new anchor, and cut between several short clips instead of asking for one long take.

Choosing the video model is a price and feature trade-off, and the gaps are wide. At the time of writing on Aitachyon, Hailuo 02 costs $0.085 per second, Kling v3 costs $0.16 per second silent and $0.32 with native audio, Wan 2.7 and Veo 3.1 Fast start around $0.19 per second, and Seedance 2.5 runs from $0.20 per second at 480p to $0.44 at 720p. For talking clips where the character speaks on camera, native audio saves a separate lipsync pass. For silent B-roll, use the cheapest model that holds the face. Our comparison of Kling v3 and Seedance 2.5 goes deeper on that choice, and the talking-head guide covers photo-plus-voice avatars specifically.

Step 3: give the character one voice and never change it

Viewers forgive a new outfit much more easily than a new voice. Pick the voice once, record its ID in the character bible, and treat it like the anchor image. You have two options:

A stock voice from a library. No consent questions, available immediately. The risk is that another account uses the same voice.
A cloned or designed voice. ElevenLabs lists Instant Voice Cloning from its $6 per month Starter plan and Professional Voice Cloning from the Creator plan. Only clone a voice you own or have written permission to use. Kapwing's guide states the same rule for faces: use real-person likeness only with ownership, rights or clear permission.

For per-call voiceover through Aitachyon, ElevenLabs costs $0.043 to $0.086 per 450 characters at the time of writing. A short reel script usually fits in one or two of those units. The full cost math is in our ElevenLabs API voiceover breakdown.

Step 4: build a posting pipeline that runs from code

The difference between an account that posts daily and one that stalls is whether a week of content can be produced in one sitting. Higgsfield lists batch production as the final step of its workflow. The version below runs from a script or from an agent such as Claude Code or Cursor, so a prompt goes in and files with their cost come out.

The weekly batch, step by step

Plan. Write seven post ideas as structured rows: format (image, carousel, reel), location, outfit, caption angle, hook line.
Assemble prompts. For each row, concatenate the identity paragraph, the visual style line and the variable details.
Generate stills. Two or three candidates per post, each referencing the approved set.
Review. A human rejects anything with drift. Keep this step manual.
Animate. Approved stills for reels go to image-to-video in short clips.
Voice. Scripts go to the locked voice ID.
Assemble and caption. Cut clips, add voice and captions, and add the disclosure (step 6).
Log. Record every file, the model it ran on and its cost next to the post it belongs to.

Wiring it to an agent

Aitachyon exposes the same models through an HTTP API (Authorization: Bearer ait_...) and a hosted MCP server. Connecting Claude Code is one command: claude mcp add --transport http aitachyon https://aitachyon.com/api/mcp --header "Authorization: Bearer ait_..."

From there the agent can take "produce this week's seven posts from plan.csv" and call the models directly. Every output carries a stable ref (img_, scn_, vo_ and so on) that a script can fetch later with GET /api/generations/{ref}, which is what makes step 8 automatic: the log is a list of refs with costs, and the files stay addressable after the session ends. The API and MCP quick start has the setup, and we documented a full working setup in automating video production with AI agents.

Two guardrails matter once an agent is spending money for you. Give the pipeline its own API key so its spend is isolated, and keep the per-key alert on so a loop that goes wrong is caught early. Failed renders are refunded automatically, so a model timeout does not silently cost you, but a prompt bug that produces 200 valid and useless images still will.

Step 5: price every post before you commit to a schedule

Know the cost of a week before you promise a daily schedule. All prices below are Aitachyon's per-call prices at the time of writing; the live grid is at aitachyon.com/api/pricing.

One-time setup

40 candidate portraits on Seedream 4.0 at $0.057 each: $2.28.
A 10-image character sheet on Nano Banana at $0.13 each, with one full redo: 20 images, $2.60.
Setup total: under $5, before any voice cloning subscription.

A realistic week (7 posts)

3 image posts, 3 candidates each on Nano Banana: 9 images at $0.13 = $1.17.
1 carousel of 5 slides, 2 candidates per slide: 10 images at $0.13 = $1.30.
3 reels, each 3 clips of 5 seconds, silent, on Hailuo 02: 45 seconds at $0.085 = $3.83.
3 voiceovers of up to 900 characters: 6 units at the top rate of $0.086 = $0.52.
Week total: about $6.82, or roughly $0.97 per post.

Swap the reels to Kling v3 with native audio so the character speaks on camera, and the same 45 seconds cost 45 x $0.32 = $14.40, which puts the week at about $17.39 before you drop the separate voiceovers. That is the real decision in this business: whether talking clips are worth roughly 3.8 times the price of silent ones for your niche. Test both and let retention decide.

How that compares with credit plans

Higgsfield prices by credits: $15 per month for 200 credits on Starter and $49 per month for 1,000 credits on Plus, with a video costing 6 to 58 credits depending on the model (vendor figures, 2026). On Plus that works out to about $0.29 to $2.84 per video, which is competitive. The trade-off is structural: a subscription is cheapest when your volume matches the plan, and per-call is cheapest when volume is uneven. We lay out the full comparison in credits versus pay-per-call, and the per-video economics of a similar setup in what a faceless channel costs to run.

Step 6: label it the way each platform requires

Most tutorials skip this step, and it is the one that gets accounts demonetised. The rules converge on one principle: realistic AI-generated people, voices and scenes must be disclosed, and platforms increasingly detect AI content on their own.

YouTube

YouTube requires disclosure when content generates photorealistic content, including realistic scenes that did not occur. A photorealistic AI influencer qualifies. You disclose in YouTube Studio during upload by selecting "Yes" under AI use in the Attributes section. Photorealistic content gets a label in the player itself. Persistent non-disclosure can lead to manual labelling, removal or suspension from the YouTube Partner Program. Fully animated or clearly fantastical content does not need the toggle. More on monetisation in what YouTube's rules say about AI content.

TikTok

TikTok requires creators to label AI-generated content that contains realistic images, audio or video, under its synthetic media policy. The manual label, launched on 19 September 2023, covers content completely generated or significantly edited by AI, and TikTok said at the time it would also test automatic labelling.

Instagram and Facebook

Meta applies an "AI info" label when it detects industry-standard AI indicators or when you self-disclose, and has labelled organic content this way since May 2024. The technical signals include provenance metadata such as C2PA Content Credentials, and automatic labelling can be over-broad, catching lightly retouched photos. For an account built on AI, Instagram added an account-level "AI creator" label, reported on 6 May 2026, shown on the bio and alongside content in Feed, Reels and Explore. Instagram told MediaPost the label does not affect algorithmic distribution.

The EU: Article 50 of the AI Act

Since 2 August 2026, Article 50 of the EU AI Act requires deployers of AI that generates deepfake image, audio or video to disclose that it is artificially generated, and requires providers to mark outputs in a machine-readable format. There is a narrower disclosure regime for evidently artistic, creative, satirical or fictional work, but a photorealistic persona selling products is a weak candidate for that exception.

The disclosure checklist

Bio states the account is AI-generated, for example "virtual creator made with AI", the wording Kapwing suggests.
Instagram "AI creator" label turned on.
YouTube altered or synthetic content toggle set to "Yes" on every realistic upload.
TikTok AI-generated label applied on every realistic post.
Do not strip provenance metadata from files to avoid labels.
Sponsored posts: "ad" or #ad in the first line of the caption and on screen in video, plus a line that the creator is virtual when that is not obvious.
No real person's face or voice without written permission.

Step 7: keep the character consistent at month three

Consistency erodes in predictable ways:

Reference creep. Someone uses last week's best post as the reference instead of the approved set. Drift compounds. Only the approved set is ever a reference.
Model switching. A new model looks better, so the reels move to it, and the face changes subtly. Re-run the character sheet on the new model and re-approve before switching.
Voice mismatch. A native-audio video model generates its own voice while voiceovers use the locked one. Pick one source of voice per format and stay with it.

Before and after: the same prompt, made reproducible

Before: "Beautiful young woman with red hair at a night market in Bangkok, cinematic, 4k."

After: "[Identity] Maya, late twenties, oval face, green eyes, copper shoulder-length hair with a side part, light freckles across the nose, small silver hoop in the left ear. [Style] Handheld phone photo, natural light, slight grain, 9:16. [Scene] Standing at a noodle stall in a Bangkok night market, holding chopsticks, laughing at something off camera, wearing a faded denim jacket. [Reference] Anchor portrait and three-quarter left from the approved set."

Three of the four blocks in the second prompt never change, which is what makes the output reproducible.

FAQ

How much does it cost to create an AI influencer?

At Aitachyon's per-call prices at the time of writing, setup costs under $5 in images and a week of seven posts with silent reels about $7. Talking reels with native audio push that to roughly $17. Higgsfield's plans start at $15 per month for 200 credits.

Do I have to disclose that my influencer is AI?

For photorealistic content, yes. YouTube requires the synthetic content toggle, TikTok a label on realistic AI content, and Meta labels detected AI content. Article 50 of the EU AI Act requires deepfake disclosure since 2 August 2026, and FTC endorsement rules cover virtual influencers in sponsored posts.

How do I keep an AI influencer's face consistent?

Use a written identity paragraph in every prompt, a fixed set of approved reference images (20 or more for Higgsfield's trained identity, up to 10 in Kapwing's builder), and generate video from approved stills with image-to-video rather than from text alone.

Can an AI influencer do brand deals?

Yes, and the same endorsement rules apply as for a human creator. Put "ad" or #ad with the endorsement itself, on screen in videos, and make clear the creator is virtual when that is not obvious to the viewer.

Does the Instagram "AI creator" label hurt reach?

Instagram told MediaPost in May 2026 that the label does not affect algorithmic distribution. A secondary source, Gradually, also reports no established reach penalty from a correct "AI info" label.

Sources

If you want to run this pipeline from a script or from Claude Code without a subscription per model, Aitachyon gives you the image, video and voice models above behind one key, billed per call from a balance that does not expire, with every file logged next to what it cost.