AI Art Pipeline

Claude orchestrates multiple AI models to generate game-ready 3D assets from text prompts -- concept art, background removal, 3D conversion, and Unreal Engine import through a single conversation.

On This Page


Overview

AgentUX lets Claude act as an art director, calling multiple AI services in sequence to turn a written game spec into 3D assets placed in your Unreal Engine level. Each asset starts as a one-line description and ends as a static mesh in the editor viewport.

Claude handles the entire pipeline: crafting prompts tuned to each model's strengths, retrying with alternative tools when quota limits hit, and importing the final assets into the UE project. The developer writes the game spec. Claude does the rest.


The 4-Step Pipeline

StepWhat HappensAI ToolOutput
1. Image Generation Claude generates concept art from text prompts, tuning prompt language per model Qwen-Image (cloud) or SDXL Turbo (local GPU) 2D concept image (PNG/WebP)
2. Background Removal Claude isolates the subject from the studio background Background Removal (Hugging Face MCP) Transparent PNG
3. 3D Conversion Claude converts the 2D image into a 3D mesh TRELLIS (Microsoft) or Shap-E (text-to-3D fallback) GLB/OBJ mesh file
4. UE Import Claude imports the mesh into the UE project and places it in the level AgentUX editor.asset + editor.actor Static mesh in the editor viewport

Each step can fail gracefully. Claude automatically falls back to alternative tools -- cloud to local GPU for image generation, TRELLIS to Shap-E for 3D conversion. The developer sees the final result, not the retries.


AI Tools in the Pipeline

Image Generation

Two tools handle concept art generation, each with different strengths:

ToolTypeStrengthLimitation
Qwen-Image Cloud (Hugging Face MCP) Understands stylistic direction from keywords like "Monument Valley art style" Daily GPU quota on free tier
SDXL Turbo Local GPU (diffusers + PyTorch) No quota limits, deterministic seeds, 4-step inference (~11s per image) Interprets style names literally -- needs explicit isolation prompts

Claude uses cloud tools first for quality, then falls back to local GPU when quotas are exhausted. Prompt language adapts automatically -- Qwen-Image gets style keywords, SDXL Turbo gets explicit product-shot instructions.

Background Removal and 3D Conversion

Background removal uses the not-lain/background-removal Hugging Face Space. 3D conversion uses Microsoft's TRELLIS model for image-to-3D, with OpenAI's Shap-E as a text-to-3D fallback for simpler geometric shapes.


Prompt Engineering

Prompt engineering is the most important skill in the pipeline. Key lessons from real asset generation:

  1. Model-specific keywords matter. Qwen-Image understands "Monument Valley art style" as an aesthetic direction. SDXL Turbo renders it as the literal national park. Claude adapts prompts to each model's training data.
  2. Explicit isolation for local models. SDXL Turbo defaults to generating scenes. "Single object, product shot, studio background, no scene, no landscape" forces it to isolate the subject.
  3. Negative prompts prevent drift. SDXL Turbo's negative prompt explicitly excludes "landscape, city, buildings, scene, environment" to keep the model focused on the object.
  4. Seeds enable reproducibility. SDXL Turbo outputs are deterministic with a fixed seed and identical parameters. Save every seed for your asset library.
  5. Cloud-first, local-fallback. Cloud MCP tools are faster and produce higher quality for stylized art. Local GPU is the safety net when quotas run out.

Try It Yourself

To use the AI art pipeline in your own project:

  1. Write a game spec describing the assets you need (name, style, rough description)
  2. Make sure Hugging Face MCP tools are available in your Claude Code session
  3. Ask Claude to generate each asset through the 4-step pipeline
  4. Claude will craft model-specific prompts, remove backgrounds, convert to 3D, and import into your UE project

Customize the prompts to match your game's art style, specify your target aesthetic, and let Claude handle the orchestration.