Do you generate consistent AI characters across images and videos?

You generate a great AI image for marketing, then the next one looks like a distant cousin of the original. Different face, different colors, different vibe. This inconsistency confuses users and weakens brand recognition, especially when you’re trying to build a recognizable character like a mascot or product hero.

Attached file: N/A

Prompt: “Generate a 3d stylized, cartoon-like yak covering his eyes with his hooves. The yak has small curved horns, shaggy light-brown hair on its head, large expressive eyes, and a friendly, slightly smiling expression. It is wearing a short-sleeved red shirt. The background is plain white, giving the image a clean, studio-like appearance. Make the yak {{ ACTION }}”

Action (left to right): “Cover his eyes with his hooves”, “sit at a table and write on a piece of paper with a pen”, “give a thumbs up”

Image model (left to right): Google Whisk, Open AI GPT Image, Google Nano Banana

❌ Figure: Bad example - The same text prompt produces inconsistent results across different image models

Step 1: Start with a single source-of-truth image

Always begin with one approved base image of the character. Ideally this image should be a character sheet, providing different views of the character so that models have more context to work with.

Use this same static image as the reference for:

• Image generation (e.g. OpenAI GPT Image, Google Nano Banana)
• Video/animation generation (e.g. OpenAI Sora, Google Veo)

Attached file: source-of-truth.png

Prompt: “Use the provided character sheet as the exact character reference. Do not change colors, facial features, proportions, or style. Make Jack the Yak {{ ACTION }}”

Action (left to right): “Cover his eyes with his hooves”, “sit at a table and write on a piece of paper with a pen”, “give a thumbs up”

Image model (left to right): Google Whisk, Open AI GPT Image, Google Nano Banana

✅ Figure: Good example - By using a source-of-truth image, you can use shorter prompts and ensure character consistency across different scenarios and models

Examples

The YakShaver team stores their source-of-truth character sheet at yakshaver.ai/media.

Figure: Character sheet showing various angles and emotions of the YakShaver mascot, Jack the Yak

The TinaCMS team stores their source-of-truth character sheet at tina.io/media

Figure: Character sheet showing various angles and emotions of the Tina mascot, Tina the Llama

Step 2: Be explicit in your prompts

AI models will improvise unless you clearly tell them not to. Include:

• Instruction to preserve the original character design
• Physical traits (species, colors, proportions)
• Clothing or accessories (or explicitly say “no clothing”)
• Mood, personality or expressions
• Camera framing and lighting

❌ Figure: Bad example - The prompt is vague and will lead to inconsistencies

✅ Figure: Good example - Explicitly instructing the model to preserve the original character

Step 3: Get approvals for generated characters

Before publishing AI-generated characters:

• All assets - Get a “test pass” from design/brand owners
• Videos & animations - Check quality with your video team to prevent choppy footage or awkward animations.