Consistent characters, mascots and products

0:00 Consistent characters and products

Making one great image of a mascot is easy. Making twenty images where she has the same face, the same scarf, the same five stripes and the same proportions? That's where most AI campaigns fall apart. By image seven, your friendly zebra has six stripes, a slightly different nose, and suddenly a brooch nobody asked for. Followers notice. Kids especially notice. In this lecture, you'll learn the consistency techniques from simplest to most advanced, how to build a character bible and a reference pack, how to keep real products accurate, and where the consent line sits for real people. By the end, you'll be able to run a recurring character or product across a whole campaign without drift.

0:51 Why consistency is hard

Why is this hard? Remember that each generation starts from new random noise. Nothing in the model remembers what your character looked like last time unless you tell it, every time, with words and pictures. And small errors compound. If a slightly wrong image becomes a reference for the next one, the next one gets more wrong. It's like a game of telephone. Each whisper changes the message a little, and by the end of the line, it's a different sentence. Consistency matters commercially too. Mascots build recognition only when they're stable, and products must match what customers actually receive. An inconsistent product image isn't a style issue. It's an accuracy problem.

1:40 Techniques, simple → advanced

Here are the techniques, from simplest to most advanced. A detailed descriptive anchor: the same precise description in every prompt. A fixed seed with small changes, useful for close variations. Character or subject references: uploading approved images. A character sheet with front, side, three-quarter and expressions, which gives both the model and your team a canonical look. Image to image from approved frames. Custom fine-tuning, like small adapter models trained on approved images, for very high consistency. And for products, compositing real product photos, which gives perfect accuracy. The newest shift is multi-reference models. Several assistant-built models now accept many reference images at once and are designed to hold identity across scenes.

2:29 The character bible

The heart of the process is the character bible. It's a one-page document with a canonical description, proportions, palette, style, expressions, wardrobe rules and do's and don'ts. For Zara the zebra: a friendly cartoon zebra, rounded features, large expressive eyes, a teal scarf, exactly five stripes across the body, head about a third of body height, flat vector style. Alongside it sits the reference pack: front, side and three-quarter views, an expressions sheet, a scarf detail, and a palette card. Here's the key idea. The bible and the reference pack are the source of truth, not the last image you generated. Every new image is checked against them before it's approved.

3:17 Worked example 1: Zara, weekly posts

Worked example one, simple. A children's education creator uses Zara in weekly storybook posts. She builds the bible and reference pack. For each new scene, she attaches the references and writes a consistency prompt: use the attached references as the only source for her design, same stripe count, same scarf, same proportions, same style. Then the new scene, Zara reading under a tree. Image seven comes back with six stripes. Instead of regenerating, she inpaints one stripe away, then checks the consistency scorecard: stripes, scarf color, head ratio, eye shape, style. All pass. Only then does image seven join the approved folder. That one rule stops the telephone game.

4:04 Worked example 2: six juice flavors

Worked example two, products, with illustrative details. An agency in Lahore produces a monthly campaign for a juice brand with six bottle flavors. Early tests show the model changes the label text, the cap color and the bottle's curve. So the agency makes a rule. The product is never generated. They photograph each bottle from three angles on a neutral background: front, three-quarter and side, with consistent lighting. The AI only generates scenes, leaving room for the bottle in the right position and angle. Then they composite the real bottle, add contact shadows, and match the grade. Result: every scene is new, and every bottle is exactly right. The client's regulatory team approves labels once, not every month.

4:55 Watch me do it: a reference pack

Watch me do it. Let me build a reference pack for a mascot from scratch. First, I create the canonical front view with the illustrator, or pick the best approved generation and clean it up by hand. Next, I generate side and three-quarter views using that front view as the reference, and I fix details manually until all three agree. Then an expressions sheet, happy, curious, surprised, thinking, all from the same references. I add a palette card with the exact brand colors. I save everything with clear names in one folder, and write the bible. Finally, I run a test: three new scenes, checked with the scorecard. If anything fails twice, I improve the references, not the prompt.

5:47 Real people: the consent line

Now real people, because consistency techniques work on faces too. A consistent AI version of a founder, an employee, an influencer or a customer needs written, specific consent. That means covering AI generation, where the images will be used, for how long, and how they can withdraw. Several platforms require labels on realistic synthetic depictions of people, and laws on likeness and personality rights vary by country. The EU AI Act's transparency rules, for example, require deployers to disclose deepfakes, including realistic images of real people, from August twenty twenty-six. When in doubt, either photograph the real person or design a clearly fictional character. Fictional characters are also easier to own and protect.

6:36 Common mistakes

Common mistakes. Starting production before the bible exists. Using drifted images as new references, which compounds errors. Describing the character differently in each prompt. Letting the model redesign the product label or the logo. Trusting a model update to keep your character the same, when new versions can interpret references differently. And skipping the scorecard when you're in a hurry. One more. Don't forget settings. If your mascot lives in a recurring place, like a classroom or a shop, make a setting reference too. Consistent places make consistent characters feel real, and they save you prompt space.

7:18 Recap and try this now

Let's recap. Consistency doesn't happen by itself, because each generation starts fresh. Use a character bible and a reference pack as the source of truth, write a consistency prompt that names the fixed attributes, check every image with a scorecard, and only approved images become references. For products, composite real photos so labels and shapes stay exact. For real people, get written consent and label realistic depictions. Try this now. Create a bible for one recurring subject, a mascot, a character or a product, with at least three reference images and a do and don't list. Then generate three new scenes and score each one before you approve it.