A content production pipeline pairing 3D asset staging with generative AI, built for on-brand beauty marketing.
Client:
LG H&H
Period:
2024
Team:
2 · Contribution 60%
Role:
3D Asset Design
AI Pipeline Design
Content Strategy
Tools:
Adobe Substance 3D
Midjourney
Rhino 3D
Photoshop
01
Chapter — System Design
The problem
Prompt-only generations
distort product shape and
shift tone between outputs.
The move
A standardized 3D stage
the generator builds on —
product shape locked.
Controlling the input,
not the prompt.
LG H&H produces marketing imagery for dozens of beauty brands at a volume traditional photography can't match. Generative AI was the obvious solution, but prompt-only workflows produced inconsistent product shapes and brand tones across outputs.
The fix was to move control upstream. The team built a library of 3D product assets and standardized Substance 3D stages — scene, lighting, camera — so the generator only fills in background and mood. Product fidelity stays consistent across runs.
I.
II.
III.
IV.
Loop — rejected outputs feed back into stage parameters, not into the prompt.
Method, not
toolchain.
With inputs standardized, the team's work shifts from prompt engineering to curation — writing segment briefs, reviewing generated frames, and making art direction calls on skin, styling, and mood.
The pipeline was documented in an internal handbook for marketers without ComfyUI experience. The method is the deliverable; the specific generator behind it is replaceable.
02
Chapter — Brand Deployment
01
ComfyUI
Base generation with segment parameters — ethnicity, age, lifestyle, values. The 3D stage defines the frame; the model generates around it.
Tool — ComfyUI / Stable Diffusion
02
Photoshop
Manual adjustments to skin, expression, hair, wardrobe, and accessories. Art direction is applied at this step.
Tool — Adobe Photoshop
03
KREA Enhancer
Realism pass. Removes the plastic-skin texture typical of AI outputs and restores photographic surface detail.
Tool — KREA AI
03
Chapter — AI Model Archiving
Project
AI Model Archiving Website
Tool
Figma
Period
Apr - Jul 2024
Team
2 · Contribution 60%
Role
UX/UI Design
System Design
AI Workflow
FILTER
민족
· Ethnicity
나이
· Age
라이프스타일
· Lifestyle
가치관
· Values
ACTIONS
+ Request new model
Submit brief
Pipeline guide ↗
F-32
20s · urban · warm
M-18
30s · minimal · cool
F-41
40s · edgy · chic
M-07
50s · classic · intelligent
F-55
20s · dreamlike · cold
F-09
30s · outdoor · bright
M-22
30s · cleancut · neutral
F-63
60s · matriarch · warm
A UI layer over
the pipeline.
Marketers at LG H&H aren't generative-AI specialists, so opening ComfyUI directly wasn't a realistic handoff. The pipeline needed a UI that matched how marketing teams already work.
The Archiving site is that layer — a library of AI models filtered by ethnicity, age, lifestyle, and values, with a request form that maps a standard brand brief ("Dr.Groot 제이몬스터즈, Saturday morning, parent with teen") to pipeline parameters.
Library filtered by
ethnicity, age, lifestyle,
and values.
Taxonomy
Request form that
maps a brand brief to
pipeline parameters.
Handoff
Built in Figma,
shipped with a usage
guide for new teams.
Delivery
04
Chapter — Midjourney Training & Lien Persona
Production time
30×
Compared to the prior human-only shoot-and-retouch workflow. Measured via production data from the Archiving site.
Internal reach
150+
Marketers registered on the AI Model Archiving site within H1 2024.
Brand outputs
3
Brand lines deployed: Dr.Groot 생화, Dr.Groot 제이몬스터즈, Physiogel — each with a dedicated model brief.
Takeaways.
Prompts describe intent but don't enforce it. A standardized 3D stage, a shared model library, and a documented handbook do — they specify camera, lighting, and the steps the team has used before. Output becomes predictable across runs.
Moving control from prompt to input was what made the pipeline reproducible across the team, not just for the people who built it.
Next Project
Bubblify
→
Sydney Hong
Seoul, KR
© Sydney Hong 2026 — Portfolio