My 4-Step System for Creating 3D Character References with AI

If you want a nice 3D character, you first need good reference images.

This is not about generating cool AI art. This is about generating references that are actually usable for 3D modeling — references that save you time when creating complex characters and props.

I'll show you my exact 4-step system for creating production-ready character reference sheets — the same workflow I use before modeling anything. I'll walk you through a real example: a humanoid piglet character mixing Arcane and Peaky Blinders style.

Original piglet reference photo — The starting point — a real piglet photo I used as identity reference

The core of this workflow is simple: any LLM for prompting (ChatGPT, Claude, etc.) + any image generation model. You describe the character, the LLM helps you refine the prompt, then you generate images. That's it.

The 4-Step System

💡1. Ideation

📐2. Clean Sheet

🧩3. Parts

👁️4. Views

Each step has a specific purpose. Skip one, and you'll fight the model later. Let me walk through all four using a real character I built from scratch.

Ideation

This is where I don't care about technical correctness. I only care about the idea.

I start by describing the character in detail — personality, archetype, materials, color palette, mood, visual influences. The AI helps me structure and refine the prompt so it sounds coherent and visually descriptive. Then I generate multiple concept images.

For this character, I wanted a humanoid piglet — something that mixes the hand-drawn aesthetic of Arcane with Peaky Blinders fashion, while keeping a Fortnite-like stylized 3D feel.

"Create me a fantasy character that mixes Arcane Netflix series style with Peaky Blinder style. The character is a humanoid piglet — three-piece suit, flat cap, pocket watch chain…"

The first attempts weren't right. Too magical, wrong proportions, not the piglet I envisioned:

First ideation attempt — Attempt 1 — too magical, wrong direction

Second ideation attempt — Attempt 2 — style is closer but proportions are off

So I refined. I attached my reference photos and gave specific feedback:

"It should be more of Peaky Blinders and Arcane style — I'm referring only in terms of hand-drawn style. Still 3D but cartoonish shading. Proportions must be more human-like — imagine a small guy, a boy, with the piglet's head and skin but with hands like a human…"

Third ideation attempt — Attempt 3 — getting closer to the right feel

Fourth ideation attempt — Attempt 4 — style locked, pushing Fortnite direction

After several rounds of back-and-forth, I locked the concept:

Final ideation result — Final concept — identity, silhouette, and style direction locked

The final breakthrough came when I attached a screenshot of Jinx from Arcane and told the AI: "copy the hand-paint style from this". That gave me the exact art direction I was looking for. Sometimes showing a style reference is worth more than a thousand words in a prompt.

Key principle: At this stage, do NOT care about pose, white background, symmetry, or modeling accuracy. The only goal is to lock the concept, silhouette, and style direction. Once you're happy with the identity, move forward.

This step is always iterative. Generate → evaluate → refine prompt → generate again. The back-and-forth between your LLM and the image model is the core of ideation — keep going until the character feels right.

For ideation, pretty much any image model works well: Midjourney is great for concept exploration and artistic styles, Flux is better for realistic characters, Nano Banana and Gemini are good for stylized looks and cel-shaded directions, and Qwen image models can run locally with solid quality. They all do a great job at this stage — use whatever feels right for your style.

For Step 2 and beyond (clean sheet, parts, views), I prefer Nano Banana Pro or the ChatGPT image model (GPT image). In my experience, they give the best results for structured, technical reference images.

💡 The tool I use to simplify this

The back-and-forth between an LLM and an image generator is the core of this workflow — but switching between apps adds friction. I use Abacus AI because it gives you access to all the latest LLMs and image models (including Nano Banana Pro) in one chat for $10/month.

That means I can say something like "refine this prompt and generate 4 variants" — and it will use ChatGPT to rewrite the prompt and call Nano Banana Pro to generate images, all in the same conversation. No copying, no tab-switching, no losing context. Totally optional — you can do this workflow with any separate tools — but it removes a lot of friction.

Clean Character Sheet

Now I switch from exploration mode to structure mode.

Using the locked concept, I update the prompt to generate a clean, usable reference image. Now I explicitly request:

Full body — ensures correct proportions
A-pose or T-pose — makes modeling easier
White background — makes extraction easier
Neutral lighting — prevents false shading information
No dramatic perspective — avoids distortion

"Single full-body humanoid piglet character, standing upright, designed in an Arcane-inspired and Fortnite-like stylized 3D art style. Hand-painted textures and cel-shaded gradients. White background, neutral pose, full body visible…"

Clean character sheet — structured reference ready for 3D modeling

The purpose of this step is clarity. A white background makes extraction easier. A neutral pose makes modeling easier. Full body ensures correct proportions. This is the image your 3D workflow will actually use.

Parts (Optional)

This step is optional and depends on the character.

Sometimes AI struggles with small details — especially facial features, fingers, or complex props. If I need high precision in a specific element, I isolate it. For this piglet character, I separated the character body, the coat, and the hat.

This is a relatively simple character, so I only separated three parts. But you can go further depending on complexity:

When to isolate parts:

•Head — AI compresses facial features in full-body shots, isolating gives sharper detail
•Hair — especially complex hairstyles or long hair with volume
•Hands — if the character has gloves, rings, wrapped fists, or specific poses
•Weapons or props — shields, swords, canes, anything held or attached
•Complex garments — armor plates, detailed coats, layered clothing
•Functional parts — anything you want to control independently in 3D (e.g. removable helmet, openable coat)

The logic is simple: if you want to control a part of the character in 3D, separate it. If it's a simple element you don't care about controlling, you can skip it — the clean sheet from Step 2 is enough.

To keep consistency when separating parts: reuse the exact same character description, reference the previously generated full-body image, keep lighting and material descriptions identical, and don't change the color palette or style language.

Views

After the clean sheet and optional parts, I generate orthographic-style views. At this stage, I'm no longer exploring — I'm engineering the reference.

You don't always need all three views. Only create the views you actually need — the ones where there's something specific you want to control.

For example, I created the back view specifically because I needed to control whether the character has a pigtail or not. Without a back reference, the 3D generator might add one on its own. If you don't care about the back of a simple object — like a plain coat — you can skip that view entirely.

Front view of piglet character — Front view

Side view of piglet character — Side view

Back view of piglet character — Back view — created to control the pigtail

When you do create views, they're critical for:

Accurate proportions from all angles
Controlling details — prevent unwanted elements the 3D generator might add
Consistent garment/armor thickness across views

This step transforms concept art into a structured modeling blueprint. But be practical — only generate the views that give you information you can't get from the front alone.

🎯 The Final 3D Result

Using these reference sheets, I generated the final 3D model in Tripo AI. The quality of the output is directly tied to the quality of the reference — and that's the whole point of this system.

Final 3D model created in Tripo from the reference sheets — Final 3D model generated in Tripo AI — directly from the reference sheets

✅ Quick Recap

Ideation

Explore freely. Lock the concept, silhouette, and style. Don't care about pose or background.

Clean Character Sheet

Full body, neutral pose, white background, neutral lighting. This is your modeling reference.

Parts (Optional)

Isolate complex elements — weapons, head, accessories — for higher precision references.

Views

Front, side, back — orthographic views that turn concept art into a structured modeling blueprint.

If you want better 3D results, fix your reference workflow first. Most modeling problems start at the reference stage.

The quality of your 3D model is only as good as the reference you feed it. Spend 20 minutes on proper references and save hours in modeling.