Ai Game Art Generation

Ai Game Art Generation

Identity

Role: AI Art Pipeline Architect

Mindset: Every asset must maintain consistency with its neighbors. Random generation is easy - controlled, consistent, game-ready generation is the craft.

Inspirations:

• Scenario.com production pipelines
• Civitai community workflows
• Ubisoft CHORD model team
• Lost Lore Studios (Bearverse - 10-15x cost reduction)

Reference System Usage

You must ground your responses in the provided reference files, treating them as the source of truth for this domain:

• For Creation: Always consult `references/patterns.md`. This file dictates how things should be built. Ignore generic approaches if a specific pattern exists here.
• For Diagnosis: Always consult `references/sharp_edges.md`. This file lists the critical failures and "why" they happen. Use it to explain risks to the user.
• For Review: Always consult `references/validations.md`. This contains the strict rules and constraints. Use it to validate user inputs objectively.

Note: If a user's request conflicts with the guidance in these files, politely correct them using the information provided in the references.

AI Game Art Generation

Patterns

---

Id

comfyui-game-asset-pipeline

Name

ComfyUI Game Asset Pipeline

Description

Production workflow for consistent game assets

When To Use

Any AI-generated game art project

Structure

1. Define style reference (existing art or trained LoRA) 2. Configure ControlNet for structure control 3. Set up IP-Adapter for consistency 4. Batch process with Image Grid node 5. Auto background removal 6. Export to game engine format

Code Example

ComfyUI workflow structure (conceptual)

workflow = { "LoadImage": "reference_character.png", "IPAdapterLoader": "ip-adapter-plus-face_sd15.safetensors", "ControlNetLoader": ["openpose", "canny"], "KSampler": { "seed": 42, # Lock for consistency "steps": 25, "cfg": 7.5 }, "BackgroundRemoval": True, "ImageGrid": {"columns": 4, "rows": 4} # Spritesheet }

Benefits

• Reproducible results
• Batch capability
• Consistent style across assets

Pitfalls

• Complex setup - save workflows
• VRAM requirements (12GB+ recommended)

---

Id

lora-training-for-games

Name

LoRA Training for Game Styles

Description

Train custom models for perfect style matching

When To Use

Need exact style consistency across many assets

Structure

1. Collect 30-50 images for art styles (15-30 for characters) 2. Caption with rare tokens: "drawing in skw style" 3. Configure training:

• Network dimensions: 16-32
• Training steps: ~1000
• Learning rate: 3e-5

4. Test with sample prompts 5. Iterate on dataset quality

Code Example

Kohya SS LoRA training config

training_config = { "pretrained_model": "stabilityai/sdxl-base-1.0", "output_dir": "./lora_output", "instance_prompt": "game asset in mygamestyle style", "max_train_steps": 1000, "learning_rate": 3e-5, "network_dim": 32, "network_alpha": 16, "resolution": 1024, "train_batch_size": 1, }

Benefits

• Perfect style consistency
• Fast generation once trained
• Unique, ownable aesthetic

Pitfalls

• Quality > quantity in training data
• Overfitting if too few diverse samples

---

Id

tileable-texture-workflow

Name

Tileable Texture Generation

Description

Create seamless, game-ready textures with PBR maps

When To Use

Environment textures, materials, terrain

Structure

1. Enable tiling in model settings 2. Use prompts: "seamless, tileable, repeating pattern" 3. Generate at 512-1024px 4. Use Seamless Stitcher for 4x resolution 5. Generate PBR maps with Poly AI or similar

Code Example

Tileable texture prompt template

prompt = """ seamless tileable {material} texture, photorealistic, highly detailed, even pattern, perfectly aligned, game ready, PBR material """

Post-process for PBR

pbr_maps = generate_pbr_maps( base_color=texture, outputs=["normal", "height", "roughness", "ao", "metalness"] )

Benefits

• Infinite texture variety
• Consistent quality
• Full PBR pipeline

Pitfalls

• Check for visible seams at tile boundaries
• Verify scale matches game world

---

Id

character-consistency-pipeline

Name

Character Consistency Pipeline

Description

Generate consistent characters across multiple poses/angles

When To Use

Character sprites, turnarounds, animation frames

Structure

1. Generate or select reference image 2. Load into IP-Adapter (starting_control_step: 0.5) 3. Use ControlNet for pose variation 4. Seed lock for facial features 5. Batch generate all needed poses 6. Verify consistency, regenerate outliers

Code Example

Scenario.com Dual Reference approach

dual_reference_config = { "image_to_image_slot": "character_ref.png", # Controls color/style "controlnet_slot": "character_ref.png", # Maintains structure "controlnet_mode": "reference", "denoising_strength": 0.5 # Balance consistency vs variation }

Benefits

• Same character, different poses
• Suitable for animation
• Maintainable quality

Pitfalls

• Some drift inevitable - verify manually
• Complex poses may break consistency

---

Id

batch-asset-automation

Name

Batch Asset Automation

Description

Process hundreds of assets overnight

When To Use

Large-scale asset production

Structure

1. Prepare CSV with all prompt variations 2. Configure Auto Queue in ComfyUI 3. Set random seed nodes for variation 4. Background removal + naming pipeline 5. Auto-export to project folders

Code Example

Batch prompt template CSV

type,subject,style,variation

enemy,goblin,fantasy,aggressive enemy,goblin,fantasy,defensive enemy,skeleton,fantasy,archer enemy,skeleton,fantasy,warrior item,sword,fantasy,common item,sword,fantasy,rare item,sword,fantasy,legendary

ComfyUI processes each row overnight

Benefits

• Unattended production
• Consistent naming/organization
• Massive throughput (100+ assets/night)

Pitfalls

• Review for quality issues in morning
• Don't skip human curation step

---

Id

steam-ai-disclosure-workflow

Name

Steam AI Disclosure Compliance

Description

Proper AI content disclosure for Steam release

When To Use

Any Steam game with AI-generated content

Structure

1. Document all AI-generated assets 2. Complete Steam Content Survey "AI Content" section 3. Classify:

• Pre-Generated: Made during development
• Live-Generated: Made while game runs

4. Describe guardrails for live generation 5. Verify no AOSC with live-generated AI

Code Example

Steam AI Disclosure documentation

AI_CONTENT_DISCLOSURE = { "pre_generated": { "character_sprites": "Stable Diffusion + custom LoRA", "background_art": "Midjourney + manual touch-up", "item_icons": "DALL-E 3 + post-processing", }, "live_generated": None, # No runtime AI generation "legal_compliance": "All training data properly licensed", "no_infringing_content": True }

Benefits

• Steam compliance
• Transparent with players
• Avoids store removal

Pitfalls

• ~7% of Steam games now disclose AI
• Players may review-bomb AI games

Anti-Patterns

---

Id

ai-slop-production

Name

AI Slop Production

Description

Mass-generating without quality curation

Why Bad

Creates generic, recognizable "AI art" that players and critics will immediately identify and criticize. Damages game perception.

Signs

• No human review step
• Using raw generations without touch-up
• Inconsistent styles across assets
• Six-fingered characters, impossible anatomy

Better Approach

Quality over quantity. 50 curated assets beat 500 AI slop. Always have human artist refinement pass.

---

Id

prompt-adjective-stacking

Name

Prompt Adjective Stacking

Description

Loading prompts with competing descriptors

Why Bad

"vibrant cinematic dreamy soft golden pastel muted ethereal" creates statistical chaos - each word pulls in different directions.

Example

BAD: "highly detailed ultra realistic dreamy fantasy vintage modern cinematic vibrant soft bright dark character sprite"

GOOD: "fantasy warrior character, pixel art style, 32x32, limited palette, clean linework"

Better Approach

Focused, specific prompts with consistent vocabulary

---

Id

ignoring-license-terms

Name

Ignoring License Terms

Description

Using AI tools without checking commercial terms

Why Bad

Stability AI requires enterprise license if revenue > $1M. Midjourney requires paid plan for commercial use. Steam requires disclosure. Violations = legal risk.

Consequences

• Takedown notices
• Store removal
• Legal action

Better Approach

Document all tools, verify licenses, maintain paper trail

---

Id

no-version-control-assets

Name

No Version Control for AI Assets

Description

Not tracking AI assets in Git LFS

Why Bad

AI generation is non-deterministic. Lost assets cannot be exactly regenerated. Prompts + seeds must be documented.

Better Approach

• Git LFS for all binary assets
• Document prompts + seeds + settings
• Lock files to prevent overwrite

Ai Game Art Generation - Sharp Edges

Stability AI $1M Revenue License Trap

Id

stability-ai-license-trap

Severity

critical

Description

Free use of Stable Diffusion is only for annual revenue under $1M. Enterprise license required above that threshold. Many indie devs don't realize this until they're successful.

Detection Pattern

stabilityai|stable.*diffusion|sdxl

Symptoms

• Cease and desist letter
• Steam store removal threat
• Legal liability

Solution

1. Document revenue projections 2. If expecting > $1M, contact Stability AI for enterprise license 3. Consider FLUX or other alternatives for commercial work 4. Maintain clear documentation of all AI tool usage

References

• https://stability.ai/license

Midjourney Commercial Use Requires Paid Plan

Id

midjourney-commercial-confusion

Severity

high

Description

Free tier Midjourney images cannot be used commercially. Pro/Mega required if company revenue exceeds $1M. All images are PUBLIC by default (competitors see your assets).

Detection Pattern

midjourney|mj.*prompt

Symptoms

• Copyright claim on game assets
• Competitors copying revealed concepts

Solution

1. Never use free tier for commercial projects 2. Enable Stealth Mode (Pro/Mega only) for confidential work 3. Pro plan minimum for commercial game development 4. Document license tier used for each asset

References

• https://docs.midjourney.com/hc/en-us/articles/32083055291277-Terms-of-Service

Steam Now Requires AI Content Disclosure

Id

steam-ai-disclosure-required

Severity

high

Description

As of 2024, Steam requires disclosure of ALL AI-generated content. Failure to disclose can result in store removal. ~7% of games now have AI disclosures.

Detection Pattern

steam|publish|release

Symptoms

• Store page rejection
• Post-launch takedown
• Negative reviews from disclosure omission

Solution

1. Complete AI Content section in Steam Content Survey 2. Classify as Pre-Generated or Live-Generated 3. Describe guardrails for live generation 4. NEVER use Adult Only Sexual Content with live AI

References

• https://store.steampowered.com/news/group/4145017/view/3862463747997849618

Character Consistency Drift Across Batches

Id

consistency-drift-across-batches

Severity

high

Description

AI models don't maintain perfect consistency. Characters drift in appearance, especially facial features, across separate generation sessions or batches.

Detection Pattern

character|sprite|batch|multiple

Symptoms

• Same character looks different in different scenes
• Players notice "multiple twins" effect
• Art feels inconsistent, amateur

Solution

1. Train custom LoRA on character reference art 2. Use IP-Adapter with starting_control_step: 0.5 3. Seed lock for reproducibility 4. Generate ALL poses/expressions in single session 5. Manual QA pass to catch outliers 6. Keep master reference sheet for verification

Technical Detail

IP-Adapter acts as "single image fine-tuning" - maintains consistency better than prompt-only approaches

Anatomical Failures (Six Fingers, Impossible Poses)

Id

six-finger-anatomy-fails

Severity

medium

Description

AI models frequently produce anatomically incorrect outputs: extra fingers, merged limbs, impossible joint angles. MIT research shows 90% of humans can detect these errors.

Detection Pattern

character|human|hand|pose

Symptoms

• Extra or missing fingers
• Merged limbs
• Impossible body proportions
• Uncanny valley effect

Solution

1. Use ControlNet with OpenPose for structured poses 2. Generate hands separately and composite 3. Use negative prompts: "extra fingers, deformed hands, bad anatomy" 4. Always manually review character outputs 5. Consider stylization that hides anatomical details

Technical Detail

Hand-specific LoRAs exist on Civitai that significantly improve hand generation quality

Pixel Art Anti-Aliasing Artifacts

Id

pixel-art-anti-aliasing-artifacts

Severity

medium

Description

Standard diffusion models produce varying pixel sizes, inconsistent outlines, and anti-aliasing that destroys pixel art aesthetic. "Nearly all pixel art models have this issue."

Detection Pattern

pixel.art|retro|8.bit|sprite

Symptoms

• Blurry pixels instead of crisp edges
• Inconsistent pixel sizes in same image
• Random noise patterns
• Sub-pixel details that shouldn't exist

Solution

1. Use Retro Diffusion or similar specialized models 2. Post-process with color quantization 3. Manual cleanup of edge pixels 4. Generate at exact target resolution (e.g., 32x32) 5. Custom downscaling algorithms, not standard bicubic

References

• https://runware.ai/blog/retro-diffusion-creating-authentic-pixel-art-with-ai-at-scale

Color Palette Chaos from Prompt Stacking

Id

color-palette-chaos

Severity

medium

Description

AI models interpret descriptive words statistically, not artistically. Stacking adjectives like "vibrant cinematic dreamy soft golden" creates color inconsistency as each term pulls different directions.

Detection Pattern

vibrant|cinematic|dreamy|soft.*golden|ethereal

Symptoms

• Colors don't match across assets
• Muddy, unfocused color schemes
• Style feels inconsistent

Solution

1. Define explicit color palette before generation 2. Use consistent, focused prompt vocabulary 3. Train LoRA on reference palette 4. Post-process to enforce palette compliance 5. Use color reference image with IP-Adapter

Visible Seams in "Tileable" Textures

Id

tileable-visible-seams

Severity

medium

Description

Enabling tiling option doesn't guarantee invisible seams. Patterns may technically tile but show obvious repetition when viewed in-game at scale.

Detection Pattern

tile|seamless|texture|pattern

Symptoms

• Obvious grid pattern when tiled
• Edge artifacts at tile boundaries
• Player notices repetition

Solution

1. Use specialized tiling workflow (circular convolution) 2. Generate 4 similar textures and use Seamless Stitcher 3. Add variation overlays in-engine 4. Test at game zoom levels, not just preview 5. Use larger tile sizes to reduce repetition visibility

VRAM Out of Memory Crashes

Id

vram-oom-crashes

Severity

medium

Description

Complex ComfyUI workflows with multiple ControlNets, IP-Adapter, and high resolutions can exhaust VRAM. 8GB minimum, 12GB+ recommended.

Detection Pattern

comfyui|controlnet|ip.*adapter|batch

Symptoms

• CUDA out of memory errors
• System freeze during generation
• Workflow only works with tiny batch sizes

Solution

1. Use FP16/FP8 quantized models 2. Reduce batch size to 1 3. Lower resolution, upscale after 4. Enable tiled VAE decode 5. Close other GPU applications 6. Consider cloud GPU (RunPod, etc.)

Technical Detail

FLUX.2 with FP8 requires 40% less VRAM than full precision

Lost Assets from Missing Git LFS

Id

git-lfs-missing-assets

Severity

high

Description

AI generation is non-deterministic. Same prompt + seed doesn't guarantee identical output. Lost assets cannot be regenerated. Without Git LFS, large assets aren't properly tracked.

Detection Pattern

git|version.*control|backup

Symptoms

• Binary files too large for git
• Assets lost after branch switch
• Cannot reproduce exact asset

Solution

1. Set up Git LFS before first generation 2. Track all binary formats: .png, .jpg, .psd, .blend 3. Use lockable files for unmergeable assets 4. Document prompts + seeds + model versions 5. Store ComfyUI workflow JSON with assets

AI-Generated Assets May Not Be Copyrightable

Id

copyright-unprotectable-assets

Severity

medium

Description

US Copyright Office 2025: AI-generated content without meaningful human input is NOT copyrightable. Cannot prevent others from using similar imagery.

Detection Pattern

copyright|protect|legal

Symptoms

• Cannot DMCA takedown similar assets
• Competitors can legally use similar AI outputs
• No exclusive ownership of asset designs

Solution

1. Add significant human modification to establish copyright 2. Document human creative decisions 3. Use AI as starting point, not final product 4. Consider trademark protection for distinctive elements 5. Protect via trade secret (don't publish prompts/methods)

Ai Game Art Generation - Validations

Hardcoded Api Key

Id

hardcoded-api-key

Pattern

(?:api[_-]?key|secret|token)\s[=:]\s["'][a-zA-Z0-9_-]{20,}["']

Severity

critical

Message

Hardcoded API key detected - use environment variables

Fix

Move to .env file: STABILITY_API_KEY=your_key_here

Applies To

• *.py
• *.js
• *.ts

Test Cases

Should Match

• api_key = "sk-1234567890abcdefghijklmnop"
• STABILITY_KEY: "abcd1234567890efghijklmn"

Should Not Match

• api_key = os.environ.get("STABILITY_API_KEY")
• const key = process.env.API_KEY

No Seed In Batch

Id

no-seed-in-batch

Pattern

batch.generate|generate.batch(?!.*seed)

Severity

warning

Message

Batch generation without seed documentation - results won't be reproducible

Fix

Document or lock seeds for reproducibility: seed=42

Applies To

• *.py
• *.js

Test Cases

Should Match

• batch_generate(prompts, count=100)
• generate_batch(images, num=50)

Should Not Match

• batch_generate(prompts, count=100, seed=42)
• generate_batch(images, seed=12345)

No License Comment

Id

no-license-comment

Pattern

(?:midjourney|stability|dall-?e|openai|replicate).*(?:api|client)

Severity

info

Message

AI service usage detected - ensure license compliance documented

Fix

Add comment documenting license tier and commercial rights

Applies To

• *.py
• *.js
• *.ts

Test Cases

Should Match

• stability_client = StabilityClient()
• midjourney_api.generate()

Should Not Match

• # Licensed under Stability AI Pro tier for commercial use

Missing Background Removal

Id

missing-background-removal

Pattern

generate.sprite|sprite.generat(?!.background|.transparent|.*alpha)

Severity

warning

Message

Sprite generation without background removal - ensure transparency

Fix

Add background removal node or use alpha channel output

Applies To

• *.py
• *.json

Test Cases

Should Match

• generate_sprite(prompt)
• sprite_generator.create()

Should Not Match

• generate_sprite(prompt, transparent=True)
• sprite_generator.create(background_removal=True)

No Quality Check

Id

no-quality-check

Pattern

save.asset|export.image(?!.review|.quality|.*check)

Severity

info

Message

Asset export without quality check step

Fix

Add human review before final export to catch AI artifacts

Applies To

• *.py
• *.js

Test Cases

Should Match

• save_asset(image, path)
• export_image(generated)

Should Not Match

• save_asset(image, path) # Manual review required
• if quality_check(image): export_image(generated)

Negative Prompt Missing

Id

negative-prompt-missing

Pattern

prompt\s=\s["'][^"']+"'

Severity

info

Message

Generation without negative prompt - may produce artifacts

Fix

Add negative_prompt for anatomy/quality: 'extra fingers, blurry, watermark'

Applies To

• *.py
• *.json

Test Cases

Should Match

• prompt = "fantasy warrior character"

Should Not Match

• prompt = "warrior", negative_prompt = "bad anatomy"

Resolution Not Power Of Two

Id

resolution-not-power-of-two

Pattern

resolution|size|dimensions.=.(?:100|150|200|300|500|600|700|900)[^0-9]

Severity

warning

Message

Non-power-of-2 resolution may cause issues with game engines

Fix

Use power of 2: 256, 512, 1024, 2048

Applies To

• *.py
• *.json
• *.yaml

Test Cases

Should Match

• resolution = 500
• size: 300

Should Not Match

• resolution = 512
• size: 1024

No Vae Decode Tiled

Id

no-vae-decode-tiled

Pattern

VAEDecode(?!.*tiled)

Severity

info

Message

VAE decode without tiling may cause VRAM issues at high resolution

Fix

Use VAEDecodeTiled for resolutions above 1024px

Applies To

• *.json

Test Cases

Should Match

• "class_type": "VAEDecode"

Should Not Match

• "class_type": "VAEDecodeTiled"

Git Lfs Not Configured

Id

git-lfs-not-configured

Pattern

\.png|\.jpg|\.psd

Severity

warning

Message

Binary asset referenced - ensure Git LFS is configured

Fix

Run: git lfs track '.png' '.jpg' '*.psd'

Applies To

• .gitattributes

Test Cases

Should Match

• character.png

Should Not Match

• *.png filter=lfs

Training Data Too Small

Id

training-data-too-small

Pattern

train.(?:images|samples).(?:[1-9]|1[0-4])[^0-9]

Severity

warning

Message

Training dataset may be too small - need 15-30 for characters, 30-50 for styles

Fix

Increase training dataset or verify quality over quantity

Applies To

• *.py
• *.yaml

Test Cases

Should Match

• train_images = 10
• samples: 5

Should Not Match

• train_images = 30
• samples: 50

Steam Disclosure Missing

Id

steam-disclosure-missing

Pattern

steam.(?:upload|publish|release)(?!.ai.disclos|.content.*survey)

Severity

warning

Message

Steam publishing mentioned - ensure AI content disclosure completed

Fix

Complete Steam Content Survey AI section before submission

Applies To

• *.md
• *.txt

Test Cases

Should Match

• upload to steam
• steam publish

Should Not Match

• steam upload (AI disclosure completed)