Stefan VaskevichHow do you create good 3D characters and assets? You need a strong idea and a good reference. So how do you get a strong idea in the first place — in the age of AI, can you just ask the model to come up with something cool?
No — and that is the single biggest mistake. That is exactly how slop is born: the moment you delegate the idea, the taste, and the judgment to the model and expect something interesting back. The AI doesn't know what's good — you do. AI is a tool. This piece is the companion to my video: how to ask the right questions, and the four rules that separate designed characters from generated ones — from concept all the way to a rigged 3D character.
I've shared a 4-step system for building good 2D references before — but that system assumes you already have an idea. This is about what happens when you don't have one yet: how to shape ideas worthy of characters no one will dare call slop.
I run this whole process inside Lovart — one canvas for the agent chat, moodboard, and image generation, so research, iteration, and the final character lock all live in one place.
Try Lovart freeThroughout, I design two characters so the rules have something to bite on: Mr. Mak, a gentleman piglet inventor — a kind tinkerer — and Mrs. Mak, his partner, an anthropomorphic supermodel piglet with full Jessica-Rabbit energy. One is cute-plush; one is glamour. The contrast is the point.
Step 0 — Before any prompt
Before a single generation, do the un-AI part of the job. Here's the actual moodboard I opened with:
- Collect a mood board. Still a valid concept. Using AI to research and gather inspiration is a fair use of the tool.
- Give the character a story, a purpose, a value. Shape it through generation, not by accident.
- Never delegate judgment or taste to AI. Slop is what AI makes when no human is reviewing.
- Create first concepts, select a few worthy ones, then apply the four rules.
Rule 1 — Silhouette + shape language
Before colour, before detail: a character has to read as a black shape against a white wall. Disney institutionalised this in the 1930s — the "10-mile silhouette" test. Shape language is emotional shorthand:
- Circlewarm, friendly, safe, naïve. (Mickey, Po, BB-8, baby Yoda.)
- Squaresolid, dependable, stubborn, slow. (Mr. Incredible, Carl Fredricksen.)
- Triangledynamic, sharp, dangerous, ambitious. (Maleficent, Scar, Syndrome.)
Applied here: Mr. Mak = circle head + square body = "kind tinkerer." Mrs. Mak = circle head + oval body on triangle heels = "showbiz drama with foundation under it." Their bodies fight; their heads match — that contrast itself is a story.
Rule 2 — Proportions
Proportion is the grammar of character design — it tells the audience what kind of person they're looking at before they read any other detail. Loomis's head-count chart (1943) is still the canonical reference:
| Heads tall | Reads as |
|---|---|
| ~3–4 | chibi / baby schema — "protect me." |
| ~4–5 | cartoon everyman — "comedic, approachable." |
| ~6 | stylized fashion / Disney heroine. |
| ~7.5 | realistic adult. |
| ~9 | heroic / superhuman / fashion plate. |
Applied here: Mr. Mak at 4.5 heads with an oversized head — a stocky inventor, the cute-on-buff contrast is the joke. Mrs. Mak at 6 heads (head plus heels do the lifting), a head-and-a-half taller than him. The visual gag is the proportion delta, not the costume.
Rule 3 — The 60 / 30 / 10 colour rule
An interior-design rule of thumb that migrated through fashion, graphic design, UI, and finally into character and costume design: 60% dominant (the colour you see first), 30% secondary (supports / contrasts), 10% accent (the spike of interest, usually near the face or hands).
Famous examples it explains:
- Mickey Mouse — black fur 60 / red shorts 30 / white gloves + yellow shoes 10.
- Spider-Man — red 60 / blue 30 / black web pattern 10.
- Jinx (Arcane) — cyan-blue 60 (inherited from Powder, her innocent self) / desaturated plum 30 / hot-pink + gold 10 (pupils, tattoos — shows up only after the transformation). Tiny surface area, huge narrative weight.
Applied here: Mr. Mak → cream 60 / leather brown 30 / copper-brass + blueprint teal 10. Mrs. Mak → stage scarlet 60 / champagne gold 30 / deep emerald 10. Their 10% accents (his teal, her emerald) rhyme on the green axis → telegraphs "couple" with no text at all.
Rule 4 — The story-iceberg
Iain McCaig's iceberg theory: 90% of a design is unseen — but felt. When the artist knows the backstory of every accessory, that knowledge leaks through their choices. Every prop tells a story.
Run this character interview before any prompt:
- What does this character want?
- What are they afraid to lose?
- What is the oldest item they wear, and why do they still wear it?
- What is the newest item, and what changed when they got it?
- What is one thing they wear that no audience would ever see — but they always know is there?
Applied here: Mr. Mak's cracked monocle is his grandfather's, deliberately unrepaired; his pocket-watch dial shows the compile-time of his current spell, not the time; the "M·B·III" on his apron was stitched by Mrs. Mak. Hers: a vintage carbon-stand mic (she insists analog "doesn't flatten the harmonics"), a hidden ruby brooch from her mother the audience never sees, and a tiny enamel pin shaped like his gauntlet tucked inside her dress strap — her secret wink from the stage.
The characters, locked
Four rules, two finished concepts. Designed — not generated.
Final step — A-pose + character sheet → 3D
Once the concept is locked through the four rules, the bridge to 3D is almost mechanical:
- Generate an A-pose on a neutral background — that's what the 3D pipeline needs.
- For props or assets with detail on multiple sides, generate multiviews (front / side / back).
- Group-export the sheet as a clean package, then run the 3D-gen pipeline (Hunyuan 3D or Tripo), clean topology in Blender, assemble, rig, and animate.
For the exact strict-front, parts, and multiview pipeline that feeds the 3D step, see my 4-step character reference workflow and the ready-made pipeline skill.
Go make something with intent
AI doesn't know what's good — you do. Use it as the tool it is: research the moodboard, run the concepts, then hold the line with silhouette, proportion, colour, and story. That's the difference between a character someone remembers and one they scroll past.
The whole process above lives in one place inside Lovart — try it free.
