HYPERTRIANGLE · AI + 3D MATERIAL PIPELINE
Making Generative AI Usable for 3D Production
HYPERTRIANGLE
OVERVIEW
Hypertriangle is an AI startup developing generative tools for 3D content creation. As generative AI continues transforming digital content production, one of the biggest bottlenecks remains creating high-quality, physically based rendering (PBR) textures that are consistent, tileable, and ready for production. The company sought to explore how generative AI could accelerate this process while giving artists greater flexibility and reducing repetitive manual work.
During my internship at Hypertriangle, I researched, designed, and developed AI-powered texture generation pipelines using ComfyUI to accelerate PBR material creation for 3D artists. My work focused on transforming fragmented AI workflows into modular production-ready tools that simplified texture generation while maintaining artist control.
ROLE
R&D Project Manager, AI Developer
TEAM
Art Team (3)
DURATION
8 Weeks
SOFTWARE
ComfyUI • JavaScript • REST API • OpenAI API • Figma
01 / PROJECT OVERVIEW
PROBLEM
Existing AI texture workflows required artists to manually combine multiple disconnected tools, making the process inconsistent and difficult to scale.
CHALLENGE
A production-ready AI workflow capable of generating PBR textures, previewing materials in real time, and exporting assets directly for 3D applications.
OUTCOME
The workflow was adopted internally and later showcased publicly by a former teammate as part of Hypertriangle's production pipeline.
Because traditional 3D production pipelines require significant time from modeling through texturing, the goal is to build a system that leverages AI to generate realistic texture data in a single workflow—applicable not only to games but also to cinematic content such as advertisements.
While AI-based texturing methods offer the advantage of generating images through simple prompt inputs, relying on a single approach can be limiting. To flexibly respond to diverse production scenarios, the following multiple approaches will be explored in parallel:
Generating texture images through simple text-based prompts
Developing high-quality texture images using Midjourney and ComfyUI
Creating textures based on real-world photographs
Enhancing and adapting textures using online texture resources
02 / SYSTEM ARCHITECTURE
Designing the material pipeline
A connected workflow turned fragmented image generation and material-prep steps into a repeatable system: generate, process, preview, adjust, and export without forcing artists to reconstruct the same asset across multiple applications.
01 / INPUT
Text prompt
Reference image
Material intent
02 / AI PROCESSING
Texture generation
Seamless processing
PBR map creation
03 / WEBGL PREVIEW
Three.js geometry
Live PBR controls
Visual validation
04 / EXPORT
PNG · JPG · ORM
Normal · Roughness
GLB asset package
03 / DESIGN + DEVELOPMENT
Working across design and development
01 / WORKFLOW DESIGN
Mapped material creation steps and identified where artists repeatedly moved files, rebuilt maps, and retested the same decisions.
02 / AI TOOL DEVELOPMENT
Built custom ComfyUI nodes with Python and JavaScript, tested emerging models, and connected generation stages into a repeatable artist-facing workflow.
03 / 3D VISUALIZATION
Developed a Three.js material viewer to inspect generated maps on multiple geometries with real-time roughness, metalness, normal, lighting, and tiling controls.
04 / TESTING + TRANSFER
Evaluated failure modes, refined parameters, and created Korean-language documentation for setup, node connections, errors, and export instructions.
≈30%
reduction in repetitive manual production work—by connecting generation, texture processing, material preview, and export into a single production workflow.
04 / TECHNICAL RESEARCH
Evaluating an emerging AI tool landscape
The work was not about choosing a single model. I compared tools by the qualities that mattered in a production workflow: output quality, texture consistency, seamlessness, processing time, controllability, integration complexity, and artist usability.
MODELS EVALUATED
Stable Diffusion XL
ControlNet · IPAdapter
Flux · LoRA workflows
Hunyuan3D · Tripo · Trellis
Sparc AI
TOOLS SELECTED FOR INTEGRATION
ComfyUI custom nodes
Python + JavaScript
Three.js + WebGL
Gemini API · OpenAI API
INTEGRATION CRITERIA
Texture consistency
Seamlessness
Control + latency
Pipeline compatibility
Artist-facing clarity
KEY DECISION
Models were treated as interchangeable capabilities—not as a product strategy. The system prioritized repeatable material quality and visible controls over a single impressive output.
Use Case 1
How the pipeline was used to produce texture for 3D artists
Renovation Guidance
05 / MATERIAL DATA
Generating physically based material maps
A realistic digital material is not defined by one image. It is a coordinated set of maps that describe how a surface responds to light—and each map needs to remain legible, organized, and predictable when an artist changes the material.
BASE COLOR
The visible surface color without lighting assumptions.
NORMAL
Adds small-scale surface depth without additional geometry.
ROUGHNESS
Controls whether reflected light appears matte, satin, or glossy.
ORM PACKING
R: Ambient Occlusion
G: Roughness
B: Metalness
Packing three grayscale maps into RGB channels reduced texture-file clutter and prepared assets for common real-time and GLB-based delivery workflows.
BREAKDOWN OF MAIN WORKFLOW
Phase 1
Phase 2
Phase 3
PART 1 TEXTURE MAKER
workflow
06 / ITERATION
What did not work at first
01 — Output quality could vary substantially between models; the workflow needed visible checkpoints rather than blind automation.
02 — Removing seams could flatten texture detail, so iteration focused on preserving surface character while improving repeatability.
03 — Flat images and rendered materials did not always agree; the real-time viewer made those differences visible before export.
04 — File paths, node communication, and dependency management required the prototype to be tested as a system—not as isolated interface screens.
DESIGNING FOR
CONTINUED USE
The tool was accompanied by Korean-language documentation for installation, workflow structure, node connections, parameters, model requirements, common errors, export instructions, and example use cases. Documentation was treated as part of the product: it made the workflow legible enough to be adopted, maintained, and extended by others.
NEXT STEPS
REFLECTION
Hypertriangle taught me to see AI as one part of a larger human workflow that must be designed, tested, and translated into tools people can actually use.
The project strengthened my interest in AI product design, human–AI interaction, creative technology, and tools that enhance rather than replace human creativity.
NEXT PROJECT