how i generate game assets with ai for my unity projects without breaking the art style every single time
AI asset generation is fast and cheap until you have six assets that all look like they came from different games. I spent four months learning how to use Midjourney, Meshy, and Unity Muse for asset generation in a way that produces a consistent visual style across a full project. These are the techniques that stopped my AI generated assets from looking like a patchwork, the mistakes that introduced visual chaos, and the workflow I now use before generating anything new.
Midjourney
Used for 2D concept art, icons, and UI elements, Standard plan $30 per month
www.midjourney.com
Meshy
3D model generation from text or image, free tier available, Pro $20 per month
www.meshy.ai
Unity Muse
Muse Texture used for tileable surface materials, $30 per month for full suite
unity.com
Marcus Webb
June 24, 2026
The problem I was solving: after two months of using AI tools to generate game assets I had a Unity project with textures that looked like painted concept art alongside textures that looked like photographs alongside assets that had a cel shaded aesthetic. Nothing matched. Players in a single playtest session used the phrase looks unfinished to describe the visual experience. The art quality was fine individually. The coherence was zero. Four months of adjusting my generation workflow later, the same playtesters describe the game as having a consistent hand-painted feel.
Why AI Asset Generation Makes Consistency Harder Not Easier
When you draw assets by hand you are the consistency mechanism. Your drawing style is your style and it appears in everything you make. AI generation has no such constraint. Every prompt produces an output based on what that prompt triggers in the model's training data. Two prompts that you think are similar can produce outputs that look nothing alike if one uses words that map to photorealism and the other uses words that map to illustration. The speed of AI generation makes this worse, not better. You can generate fifty inconsistent assets in an hour that would have taken two weeks to make by hand. Speed without a consistency system produces a mess faster.
The Style Reference Document I Create Before Generating Anything
# My art style reference document for AI generation
# Created once per project, used in every generation prompt
## Core style description (under 30 words)
Hand-painted RPG environment art, warm earthy palette,
soft shadows without hard outlines, slightly stylized
proportions, reminiscent of early 2000s PC RPG aesthetic.
## Color palette (5 values maximum)
Primary: muted terracotta #B5651D
Secondary: deep forest green #355E3B
Neutral: aged parchment #F5DEB3
Shadow: warm dark brown #3B2314
Accent: faded gold #C5A028
## What this style is NOT
Not photorealistic. Not cel shaded with black outlines.
Not high saturation. Not watercolor. Not pixel art.
## Reference keywords for every Midjourney prompt
[paste these into every generation prompt]
hand painted, rpg concept art, warm earthy tones,
slight stylization, soft ambient lighting, no outlines,
early PC RPG aesthetic, muted palette
## Reference keywords to AVOID
photorealistic, 3D render, watercolor, cel shaded,
neon, vibrant, high contrast, anime style
## My first generated asset
[Paste the Midjourney job ID or image link of the first
asset you were happy with. Use the --sref flag in all
subsequent generations to reference this image's style]
# HOW I USE THIS
# Before generating any new asset I open this document.
# My Midjourney prompt always ends with my reference keywords.
# My Muse Texture prompts describe the surface then append
# the style description at the end.
# My Meshy texture prompts use the color palette values
# and style keywords for the material generation step.
# This document lives in a Notion page called
# [ProjectName] Art Style Reference.
# If the document is not open I do not generate.How I Use Each Tool in the Asset Pipeline
- Midjourney Standard at $30 per month for 2D assets: UI icons, inventory item illustrations, character portrait references, and concept art for 3D modelers to work from. I use the sref flag to reference my approved first generation and the style keywords from my reference document in every prompt. Consistency improved dramatically once I started doing both of these things.
- Meshy free tier moving to Pro at $20 per month for 3D props: Generates rough 3D prop geometry from text descriptions. The output is never final but it reaches playable placeholder quality fast. For stylized games the texture generation step in Meshy is important for matching your palette. I paste my color values into the texture generation prompt.
- Unity Muse Texture at $30 per month as part of the Muse suite for surface materials: Tileable textures for environment surfaces. Stone walls, wooden floors, dirt paths. I append my full style description to every Muse Texture prompt. Output still needs review but the style drift is manageable when the prompt includes the style constraints.
- Poly Haven free tier for base HDRIs and some props: Does not replace AI generation but Poly Haven CC0 assets are free and professionally made. I use them for HDRI lighting and for any prop type that exists in their library. Mixing AI generation with Poly Haven requires color grading the Poly Haven assets to match my palette.
Mistakes That Introduced Visual Chaos
- Generating assets from memory instead of from the reference document: Three times I generated an asset thinking I remembered the style keywords well enough. All three times the result had a subtle style drift that I did not notice until I placed the asset next to existing ones in the scene. The document is not optional.
- Mixing AI generated assets with free pack assets without color grading: I added a Unity Asset Store environment pack to my project in month two because it saved time. The pack had a higher saturation and cooler color temperature than my AI generated assets. The level looked like two different games had been stitched together. Spent two days in Blender color matching the pack assets to my palette.
- Using the wrong model for stylized output: Midjourney version 6 defaults to photorealism unless you push against it with style keywords. My first month of generations had too much photorealistic detail because I was not including enough stylization instructions. Version 6.1 with the sref flag and explicit style keywords produces much more consistent stylized results.
- Generating at low resolution and scaling up: Early texture generations at 512 pixels looked fine in the thumbnail and lost all the painted texture detail when applied to a large surface in Unity. Now I generate Midjourney reference art at the highest resolution and export Muse textures at 2048 minimum.
What the Workflow Looks Like in Practice
- Open the art style reference document before generating anything.
- For a new 3D prop: generate the mesh in Meshy first, then use the Meshy texture generation with my color values to paint the surface, then export as GLB and import to Blender for cleanup, then export as FBX into Unity.
- For a new UI element or icon: generate in Midjourney with the reference document keywords plus the sref flag pointing to my first approved generation. Three variations, pick one, adjust colors in Photoshop to exactly match my palette hex values.
- For environment surfaces: generate in Muse Texture with the style description appended to every prompt. Review at multiple UV scales in Unity before committing.
- Before adding any new asset to the project: place it next to three existing assets in a blank Unity scene. Check whether it reads as the same game. If it does not, adjust before it enters the main project.
Final Thoughts
Four months in, AI generated assets look like they belong in my game because I built a system to keep them consistent before generating them, not after. The style reference document, the review step before adding to the main project, and the palette enforcement in every prompt are the three habits that changed the result. The tools are fast. Consistency is the constraint they do not provide on their own. You have to build the system.