Generative AI in creating content with a brand character
Why a character 'floats' from generation to generation
Most teams start with a textual description: 'girl 25 years old, red hair, green eyes, casual style'. The neural network honestly fulfills the request — and each time draws a new person who simply fits these parameters. The problem is not with the model, but with the approach: a textual description is not an identifier, but a range of probabilities.
A working solution is built on three levels simultaneously, not on just one of them:
- Visual references — a fixed set of images of the character from different angles, which are fed to the model as a standard, not as inspiration.
- ID control of the character — a technology that binds the face and key features to a specific dataset, not to a textual prompt.
- Character rig AI — fixation of the skeleton, proportions and facial expressions, which keeps the pose and expression within the boundaries set during training.
Remove any one of the three levels — and you get identity drift: the character is recognizable on the first image and becomes a stranger on the fiftieth. This is exactly what happens in 90% of attempts to create an AI character on your own — the team bets on prompt engineering where infrastructure is needed.
System prompt: 8 components without which the brand voice blurs
A prompt like 'write as a young friendly brand' doesn't work beyond one generation. To make AI reproduce a character's style and tone rather than a command, the system prompt must cover eight mandatory areas.
- Role — who exactly is speaking and on whose behalf.
- Tone — specific adjectives and antonyms (‘friendly, but not sycophantic’).
- Values and prohibitions — what the character will never say or do.
- Vocabulary — brand-specific words and phrases to avoid or, conversely, to use.
- Response structure — length, rhythm, presence of emojis, form of address.
- Examples — 2–3 reference answers of ‘how it should be’ and at least one counterexample of ‘how it should not be’.
- Behavior under uncertainty — what the character does when they don’t know the answer.
- Quality requirements — criteria by which the editor checks the result.
Without the last two items, AI starts hallucinating with a confident tone of the character — this is more dangerous than a simple mistake, because the audience perceives the fabrication as the brand's voice.
A beautiful image or a media asset: what is the difference
An aesthetic AI character and a brand character — different products, even if visually they match in the first month. The difference becomes apparent not at the start, but after the tenth, twentieth, hundredth post.
| Criterion | Just a beautiful AI character | A brand media asset |
|---|---|---|
| Consistency | Holds for 3–5 generations, then starts to drift | Fixed via control ID and character rig |
| Voice | Changes depending on the copywriter’s prompt | Fixed via a system prompt with 8 components |
| Scalability | Breaks when transitioning to a series of content | Works with 50+ posts per day |
| Usability when changing positioning | Requires complete rebuild from scratch | Re-trains through an adaptive layer |
| Economics | Saves designer's time once | Reduces content cost systematically, affects CTR and revenue |
The difference in numbers is noticeable: companies with recognizable and retained unified voice tone show revenue growth by 23–33% compared to those who don't maintain ToV. A pretty picture doesn't give this effect — it's not reproducible at scale.
Case: how an AI character changed the content economy on a marketplace
A natural cosmetics brand on Wildberries transitioned from a classic production to an AI character and achieved a result rarely mentioned in cases about 'convenient tools'. Content cost dropped from 1 500 ₽ to 119 ₽ per post. The company now produces 50 posts per day instead of two — that is, not just cheaper, but ten times faster.
More illustrative is another indicator: CTR of product cards rose from 2,1% to 5,5%, while the cost per view decreased from 1,2 ₽ to 0,03 ₽. This happened not because AI images are more beautiful than photo sessions, but because the character remained recognizable in each of the fifty publications per day — without this condition, the volume of content would have had the opposite effect: scattered visuals blur the product card rather than strengthening it.
Typical mistakes when training an AI character
Most failures are repeated from brand to brand, and almost all of them are managerial rather than technical.
- Copying other people's ready-made prompt templates instead of writing original texts based on a real brandbook.
- Overloading the prompt with excessive instructions — the model begins to ignore part of the requirements when there are too many.
- Lack of retraining when positioning changes — the brand changes, but the character continues to speak with an old voice.
- Mixing the character, object, and style in one generation instead of separating these entities — hence the 'floating' facial features when changing the background or clothing.
- Ignoring the risk of data leakage: according to studies by Cyberhaven and LayerX, 77% of employees paste data into GenAI prompts, while less than 1% of employees are responsible for 80% of data leak incidents, including internal documents — working with a brandbook and character guidelines presents a direct risk to the brand.
There is also a reputational risk on the audience side. In a WPP Media study, only 42% of people were able to correctly identify an AI voice in a single sentence and 47% — in a full audio advertisement. Being aware that you are looking at AI content reduces the perceived authenticity — especially in emotional messages. This is not an argument against AI characters, but against making them indistinguishable from real people without transparency.
End-to-end process: how not to lose unity when scaling
Translating a brand book into parameters for a neural network is a separate task, and it must be solved before generating the first image, not after. The working model for managing a character's voice when scaling is built from five levels.
- Voice core — immutable characteristics: name, values, visual code. These change only during a full rebranding.
- Adaptive layer — variations in tone for different channels (stories vs product card vs email).
- Instruction Management — a library of prompts, versioned like code, not rewritten from scratch by each employee.
- Editorial Oversight — a person checks every batch of content before publication, not just the first one.
- Ethics Level — rules for disclosing that the character is AI-generated, and boundaries of acceptable topics.
Without this five-level model, the brand voice blurs predictably: first, small tonal details are lost, then large value positions, and in the end, the content starts to look as if the brand has 'dissociative identity disorder' — different posts contradict each other in their manner of speaking. Full automation without editorial oversight accelerates exactly this process, not just production. The balance is not achieved by reducing human involvement, but by moving the human from generation to quality control — this is the only point that is not yet replaced by ID control or character rig.
Why a character 'floats' from generation to generation
Most teams start with a textual description: 'girl 25 years old, red hair, green eyes, casual style'. The neural network honestly fulfills the request — and each time draws a new person who simply fits these parameters. The problem is not with the model, but with the approach: a textual description is not an identifier, but a range of probabilities.
A working solution is built on three levels simultaneously, not on just one of them:
- Visual references — a fixed set of images of the character from different angles, which are fed to the model as a standard, not as inspiration.
- ID control of the character — a technology that binds the face and key features to a specific dataset, not to a textual prompt.
- Character rig AI — fixation of the skeleton, proportions and facial expressions, which keeps the pose and expression within the boundaries set during training.
Remove any one of the three levels — and you get identity drift: the character is recognizable on the first image and becomes a stranger on the fiftieth. This is exactly what happens in 90% of attempts to create an AI character on your own — the team bets on prompt engineering where infrastructure is needed.
System prompt: 8 components without which the brand voice blurs
A prompt like 'write as a young friendly brand' doesn't work beyond one generation. To make AI reproduce a character's style and tone rather than a command, the system prompt must cover eight mandatory areas.
- Role — who exactly is speaking and on whose behalf.
- Tone — specific adjectives and antonyms (‘friendly, but not sycophantic’).
- Values and prohibitions — what the character will never say or do.
- Vocabulary — brand-specific words and phrases to avoid or, conversely, to use.
- Response structure — length, rhythm, presence of emojis, form of address.
- Examples — 2–3 reference answers of ‘how it should be’ and at least one counterexample of ‘how it should not be’.
- Behavior under uncertainty — what the character does when they don’t know the answer.
- Quality requirements — criteria by which the editor checks the result.
Without the last two items, AI starts hallucinating with a confident tone of the character — this is more dangerous than a simple mistake, because the audience perceives the fabrication as the brand's voice.
A beautiful image or a media asset: what is the difference
An aesthetic AI character and a brand character — different products, even if visually they match in the first month. The difference becomes apparent not at the start, but after the tenth, twentieth, hundredth post.
| Criterion | Just a beautiful AI character | A brand media asset |
|---|---|---|
| Consistency | Holds for 3–5 generations, then starts to drift | Fixed via control ID and character rig |
| Voice | Changes depending on the copywriter’s prompt | Fixed via a system prompt with 8 components |
| Scalability | Breaks when transitioning to a series of content | Works with 50+ posts per day |
| Usability when changing positioning | Requires complete rebuild from scratch | Re-trains through an adaptive layer |
| Economics | Saves designer's time once | Reduces content cost systematically, affects CTR and revenue |
The difference in numbers is noticeable: companies with recognizable and retained unified voice tone show revenue growth by 23–33% compared to those who don't maintain ToV. A pretty picture doesn't give this effect — it's not reproducible at scale.
Case: how an AI character changed the content economy on a marketplace
A natural cosmetics brand on Wildberries transitioned from a classic production to an AI character and achieved a result rarely mentioned in cases about 'convenient tools'. Content cost dropped from 1 500 ₽ to 119 ₽ per post. The company now produces 50 posts per day instead of two — that is, not just cheaper, but ten times faster.
More illustrative is another indicator: CTR of product cards rose from 2,1% to 5,5%, while the cost per view decreased from 1,2 ₽ to 0,03 ₽. This happened not because AI images are more beautiful than photo sessions, but because the character remained recognizable in each of the fifty publications per day — without this condition, the volume of content would have had the opposite effect: scattered visuals blur the product card rather than strengthening it.
Typical mistakes when training an AI character
Most failures are repeated from brand to brand, and almost all of them are managerial rather than technical.
- Copying other people's ready-made prompt templates instead of writing original texts based on a real brandbook.
- Overloading the prompt with excessive instructions — the model begins to ignore part of the requirements when there are too many.
- Lack of retraining when positioning changes — the brand changes, but the character continues to speak with an old voice.
- Mixing the character, object, and style in one generation instead of separating these entities — hence the 'floating' facial features when changing the background or clothing.
- Ignoring the risk of data leakage: according to studies by Cyberhaven and LayerX, 77% of employees paste data into GenAI prompts, while less than 1% of employees are responsible for 80% of data leak incidents, including internal documents — working with a brandbook and character guidelines presents a direct risk to the brand.
There is also a reputational risk on the audience side. In a WPP Media study, only 42% of people were able to correctly identify an AI voice in a single sentence and 47% — in a full audio advertisement. Being aware that you are looking at AI content reduces the perceived authenticity — especially in emotional messages. This is not an argument against AI characters, but against making them indistinguishable from real people without transparency.
End-to-end process: how not to lose unity when scaling
Translating a brand book into parameters for a neural network is a separate task, and it must be solved before generating the first image, not after. The working model for managing a character's voice when scaling is built from five levels.
- Voice core — immutable characteristics: name, values, visual code. These change only during a full rebranding.
- Adaptive layer — variations in tone for different channels (stories vs product card vs email).
- Instruction Management — a library of prompts, versioned like code, not rewritten from scratch by each employee.
- Editorial Oversight — a person checks every batch of content before publication, not just the first one.
- Ethics Level — rules for disclosing that the character is AI-generated, and boundaries of acceptable topics.
Without this five-level model, the brand voice blurs predictably: first, small tonal details are lost, then large value positions, and in the end, the content starts to look as if the brand has 'dissociative identity disorder' — different posts contradict each other in their manner of speaking. Full automation without editorial oversight accelerates exactly this process, not just production. The balance is not achieved by reducing human involvement, but by moving the human from generation to quality control — this is the only point that is not yet replaced by ID control or character rig.
Recent articles
MegaFon, BMW, VTB and more 240+ companies. See how it looks in the showcase reel:
Our branding agency is approached to order:
Branding agency services
