I Let Nano Banana Redesign My Face 5 Times | How I Found My Brand’s Visual Style
What testing five art styles taught me about building a consistent visual identity—through systems, rules and the right mix of AI tools. [Prompt Review]
This post is part 1 of a series on Branding with AI. Read next posts by clicking here.
Every AI creator hits a point where generic visuals stop feeling like a brand and start feeling like a placeholder. Mine arrived the day I saw Build to Launch’s stunning 3D Pixar-style thumbnails — and the beautiful visual systems emerging across newsletters like The AI Break, ROBOTS ATE MY HOMEWORK, The AI Rabbit Hole. I really appreciate their attention to visual storytelling. I looked at my latest “AI photo” and thought: This could be anyone’s post.

Reading Jenny Ouyang’s Substack post on consistent hero images lit the spark. One test later, I had an entire illustration workflow, five versions of Pinkie (my online persona), and a very clear winner.
Before we dive in, here’s what you’ll learn in this post:
How to build a system for consistent AI images that supports your brand
What happened when I tested five illustration styles on the same character
Why systems thinking beats prompting to create scalable visual identity
Jenny’s Consistent AI Hero Image System
Creator: Jenny Ouyang
Source: “How I Create Consistent Hero Images, And Why I Haven’t Switched to NanoBanana”
TL;DR: Jenny’s post doesn’t give “the perfect prompt.” Instead, she shares a process for consistent character illustration:
Build a base library of character poses (front, side, back, happy, neutral, talking, walking)
Store them in one workspace
Feed reference images to ChatGPT to generate the style-specific prompt
Reuse system prompts and brand rules
Her core insight: Set up your system correctly, and any prompt becomes consistent.
This idea became the backbone of my own experiment: testing five illustration styles for Pinkie to find the one that felt like AI Meets Girlboss.
My Test: What Happens When You Give Your Face to 5 Art Styles?
Before committing to a brand aesthetic, I wanted to see how Pinkie transformed across styles. I went full teenage-graphic-designer mode (shoutout to my Avril Lavigne fan-site years).
Here’s my exact workflow:
Step 0: Pick the base face.
This is my original profile photo—Pinkie, Founder of AI Meets Girlboss. She doesn’t know it yet, but she’s about to get a make-over, AI’s going to take over and turn her into an illustration.
Step 1: Gather five illustration art styles.
After exploring Google + Pinterest, I chose:
Comic Book
3D Pixar Cartoon
Fashion Illustration
Caricature
Retro Poster
Step 2: Turn references into “style rules.”
I fed each reference image into ChatGPT using my Blind Illustrator Prompt:
Analyze this [style name] illustration style as if describing it to a blind illustrator. Identify the exact visual rules: shapes, proportions, textures, color palette, lighting, line quality, shading, character features, poses, composition, and mood.
[Attach reference image]This gave me a structured breakdown of each aesthetic.
Step 3: Turn style rules into a transformation prompt.
For this step, I built The Style Transformation Prompt and asked ChatGPT:
Use these rules to write a prompt beginning with: “Transform the person in the input image into a [style] illustration.” Include line work, shading, textures, colors, facial expression logic, pose, and atmosphere. Write it as a complete, polished prompt that an image model can use to recreate the style consistently.
[Insert rules here]This created a reusable style prompt for each variation.
Step 4: Generate the illustrations.
ChatGPT kept changing Pinkie’s face like she was in a witness protection program, so after a few tries, I moved to Nano Banana. I tested each style on the same base photo: my Pinkie persona.

Step 5: Pick the style that feels like AI Meets Girlboss.
This part needed human judgment. Emotion plays a role in brand identity.
Here are the five final styles I created with NanoBanana:
⬆️ I’m curious — really! Before you scroll further, please do tell me which style is your favorite! ⬆️
Step 6: Psst… I Had a Favorite, So I Developed Her Further
The moment Fashion Illustration Pinkie appeared, I knew she was the one. She had that feminine confidence, that “I belong here” energy. She felt like the AI Meets Girlboss brand.
And yes, I’m fully aware I’m choosing a feminine, couture-forward visual identity on a platform where the best-selling rows are dominated by men in a masculine tech environment. That’s exactly why I chose her. If Pinkie’s going to show up, she’s going to show up in style.
But the first versions from NanoBanana, though good, still felt “half-finished”. I wanted more polish. So I used ChatGPT 5.1—not for generation, but for refining the style rules:
consistent proportions
controlled texture
clarity on what can stay “unfinished” (background washes)
clarity on what must stay consistent (face, posture, palette)
Once the rules were locked, I returned to NanoBanana to generate consistent poses.
Step 7: Mutations + Image Library
With style rules in place, I built a full image library: standing, walking, talking, presenting, thinking, close-ups, seasonal outfits etc. You get the drill. I really wanted to make you proud Jenny Ouyang + it was a HELL lot of FUN.
NanoBanana kept Pinkie perfectly consistent across all of them. Pinkie’s visual identity system — consistent, stylized, repeatable, mutation-ready was born.
I even rebranded one of my older post covers to test the identity in the wild:
What I Learned From This Whole Experiment
1. What was consistent for Jenny was not consistent for me.
Jenny gets incredible consistency with ChatGPT 4o. I did not. No matter how many reference images I tried, ChatGPT kept reinventing Pinkie’s face, outfit, or proportions. But ChatGPT was exceptional at:
analyzing styles
refining visual rules
producing high-end prompts
On that, Jenny and I completely agree: ChatGPT has the nicer finish. It’s just not my consistency engine.
2. NanoBanana is unmatched for character consistency.
The moment I moved to NanoBanana, image consistency snapped into place.
Pinkie finally had the same face, the same colors, the same identity across dozens of poses.
NanoBanana is simply built for character continuity, and for my workflow, that’s priceless.
3. Systems > prompts.
Prompts are one-time instructions. Systems are infrastructure. Jenny’s biggest lesson was to build a system that remembers your character and your style rules.
The Lesson That Took Me 50+ Images to Understand
A huge thank you to Jenny Ouyang for sharing her method!
So here’s the tea, Queen: Jenny’s workflow is brilliant, but she’s Jenny, and you’re you. Her perfect setup wasn’t a plug-and-play solution for my world, and yours might look different again.
AI isn’t a recipe. You learn its quirks, you figure out what works for your character and your brand. You try things, you break things, and somehow you end up exactly where you needed to be.
This is why learning AI matters. Not to copy someone else’s blueprint—but to understand how to build your own. Because once you know how to work through the obstacles, you can always, always find your finish line.
Later Girlbosses,
Pinkie 🩷🦩
P.S.
Next week: the full behind-the-scenes on how I create 50+ mutations without breaking Pinkie’s face? Pinkie Promise. 😉
Want the Girlboss AI Starter Kit? Five elite prompts for women who want to use AI strategically. Subscribe and I’ll send it your way.
Growth comes from practice
I’m committed to actually using these systems — not just testing them. So here are a few well-earned shoutouts to the Substackers who helped shape this edition:
Thank you to Claudia Faith for the One Prompt System Outline, which helped shape this newsletter.
And a thank you to Mia Kiraki 🎭 whose Make AI Argue With Itself Prompt helped me choose today’s title.








![The AI Jargon Hangover [#1 This Week in AI Tea]](https://substackcdn.com/image/fetch/$s_!MsaA!,w_140,h_140,c_fill,f_auto,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f300306-6fb3-4e17-8749-4118279cc643_1600x896.png)
Love this, I'm on the journey too, and you're ahead of me sorting it all out. Yay!
So next post is more on this?
So I used ChatGPT 5.1—not for generation, but for refining the style rules:
consistent proportions
controlled texture
clarity on what can stay “unfinished” (background washes)
clarity on what must stay consistent (face, posture, palette)
Love this! You just gave me so much insight and solved one of my biggest Substack problems.