Generative AI / Product Visualization / Workflow Design
Project
LG Household & Health Care · Team of 3
Period
March–September 2024
Role
AI Design Intern
3D/AI workflow
Content and guides
Tools
Midjourney
Substance 3D Stager
Stable Diffusion
ComfyUI
Photoshop
Production Efficiency
Compared to the traditional person-based shoot-and-retouch workflow. Measured from the archiving site's production data.
Revenue Efficiency
Marketers registered on the AI Model Archiving site in the first half of 2024.
Brand Outputs
Brand lines deployed: Dr.Groot Fresh Flower, Dr.Groot J.Monsters, Physiogel — each with a dedicated model brief.
01 — The Challenge
Changes to the product led to repeated rework.
Beauty content needs to show the product accurately as well as create the right mood. AI offered different visual directions, but distorted products still needed retouching.
Repeated corrections led me to review the full process, including the work after generation.
02 — 3D and AI roles
Use 3D for the product and AI for the setting.
Beauty content needs to show the product accurately as well as create the right mood. AI offered different visual directions, but distorted products still needed retouching.
Repeated corrections led me to review the full process, including the work after generation.
01
Product silhouette, proportions, and packaging
3D assets as a consistent reference
02
Background, mood, and concept variations
AI-assisted visual exploration
03
Brand fit and product accuracy
Designer selection, review, and retouching
03 — Reusable assets
Prepare product assets that can be used across scenes.
I proposed a 3D product asset library and helped build it. Reusing the same asset would give different scenes and concepts a consistent product reference.
04 — Scene setup
Set product placement and lighting in Stager.
I proposed using Substance 3D Stager and worked on product placement, cameras, and lighting. Adjusting these scenes helped establish render settings for the next stage of AI production.
05 — Production workflow
Beyond One-Off Generation: Establishing a Sustainable 'AI Model Archiving System'
Beauty marketing content requires precise tailoring to match target demographics and brand aesthetics. Instead of relying on repetitive, one-off image generation, I proposed and helped build a dedicated AI Beauty Model Library, directly contributing to its UX/UI design and asset management.
As demonstrated in the top examples, this system allows for flexible visual expansion using the same base models and 3D assets—ranging from a dreamlike underwater floral mood targeting millennials and Gen Z, to natural, everyday lifestyle shots for families. By creating an intuitive interface where marketers can seamlessly browse required personas and request specific content, I successfully scaled individual technical workflows into an efficient, team-wide asset.
The library profiles each model by ethnicity, age, lifestyle, and values, and runs as an internal model website that teams across LG H&H can use. For requesters new to generative AI, I also wrote a request guideline on what to specify and how. Once it was in place, production time dropped to one-thirtieth of the previous process. About 150 models were registered as of the first half of 2024, and the team's verdict was that content could now be planned around the AI output itself.
01
AI Beauty Model Library
AI 뷰티 모델 라이브러리
An internal model website where marketers filter virtual models by ethnicity, age, lifestyle, and values to plan upcoming content.
Filters · 필터
Ethnicity · 인종
Age · 연령
Lifestyle · 라이프스타일
Values · 가치관
Actions · 액션
+ Request New Model · 새 모델 요청
Submit Brief · 브리프 제출
Pipeline Guide ↗ · 파이프라인 가이드
150+ models · scroll to browse ↓ · 150개 이상의 모델 · 스크롤해서 보기
02
Beauty campaign mockups tailored to specific target demographics
특정 타깃층에 맞춘 뷰티 캠페인 목업
(e.g., women in their 20s–30s, families) and brand guidelines, created by integrating Generative AI with 3D assets.
Dr.Groot · Fresh Flower Edition
Underwater floral, targeting 20s–30s.
Roses in water, a cool palette, low-contrast lighting. A dreamlike beauty visual with editorial mood.
Dr.Groot · J.Monsters
A parent and teenage child
roses in water, a cool palette, low-contrast lighting. A dreamlike beauty visual with editorial mood.
Physiogel
Parent and child, a neutral palette.
parent–child pairing, flat lighting, calm wardrobe, restrained color. Reference: a clinical brand tone.
DAISO
One archived model, four request sheets
아카이브 모델 한 명, 요청서 네 장
For The Face Shop skincare sold at Daiso, the brand team sent one request sheet per cut, and every sheet named the same archived AI model, WA033. Each sheet fixed the pose, hair, makeup and wardrobe against a reference photo. My job was to keep WA033's face identical across every cut and change only what each sheet asked for. Backgrounds stayed plain on purpose, so the SI team could recolor them later to match each store's VMD layout. The desk shot went out in two versions, with and without cream texture on the cheek.
다이소에서 판매되는 더페이스샵 기초 라인을 위해 브랜드팀이 컷마다 요청서를 보냈고, 모든 요청서가 아카이브에 등록된 같은 AI 모델 WA033을 지정했습니다. 요청서마다 포즈·헤어·메이크업·의상이 레퍼런스 사진으로 정해져 있었습니다. 제 역할은 모든 컷에서 WA033의 얼굴을 똑같이 유지하면서 요청서가 바꾸라고 한 요소만 바꾸는 것이었습니다. 배경은 SI팀이 매장별 VMD 시안에 맞춰 나중에 색을 바꿀 수 있도록 일부러 단색으로 두었고, 책상 컷은 볼에 크림 제형이 있는 버전과 없는 버전 두 가지로 전달했습니다.
01 · Brief · 요청서
One sheet per cut · 컷마다 요청서 한 장
02 · Base cut · 기본 컷
WA033, posed to the reference · 레퍼런스 포즈에 맞춘 WA033
03 · Variant · 변형
Same cut, cream texture added · 같은 컷에 크림 제형 추가
04 · Next sheet · 다음 요청서
Same face, new hair and wardrobe · 같은 얼굴, 새 헤어와 의상
The Face Shop · Daiso
Model WA033, four cuts from four sheets.
Pose, hair, makeup and wardrobe changed with every sheet; the face did not. Plain backgrounds left room for the SI team to recolor each cut for in-store VMD.
Date · 날짜
2024.06–08 · 2024년 6–8월
Tools · 툴
ComfyUI — WA033 generation, pose from the reference · 레퍼런스 포즈로 WA033 생성
Photoshop — hair, hands, wardrobe color, cream texture · 헤어·손·의상 컬러·크림 제형 보정
Enhancer — resolution and eye detail · 화질과 눈동자 디테일 개선
Sooryehan (수려한)
From brief to the store shelf
브리프에서 매장 진열대까지
01 · Brief · 브리프
A composed Chinese woman, 30s
02 · Revision · 수정
Changed to a Korean woman with a Chinese feel
03 · Detail · 디테일
Skin texture and hair detail
Sooryehan
A Korean model with a Chinese feel, 30s.
The brief began as a composed Chinese woman in her 30s, then changed to a Korean model with a Chinese feel. Skin tone, expression, wardrobe, hair, and accessories were reworked, and skin texture was refined before the model ran as an in-store poster.
Date
2024.05
Tools
ComfyUI — base model generation
Photoshop — wardrobe, hair, accessories, proportions
KREA Enhancer — realistic skin, removing the plastic Midjourney texture
AVON
A scalp shot with almost no references
Avon had one composition in mind: a dropper applying ampoule straight to the scalp. Almost every reference showed ampoule going on the face, not the scalp, so generating the shot in a single image kept failing. I split the problem instead. The AI beauty model's face and composition were built on their own, the dropper was generated separately, and the two were composited to read as one photograph. Avon approved the result and used it on its official product page.
01 · Model · 모델
Face and composition, built on their own
02 · Dropper · 스포이드
The dropper, generated separately
03 · Composite · 합성
Merged into one photograph
04 · In use · 실사용
Live on Avon
Client
Avon
Live
View on avon.com ↗
07 — Guides and training
Turn one person's method into the team's method.
Knowing a tool well was not enough if only one person on the team could use it. When colleagues asked how the beauty visuals were made, I turned my own trial and error into an internal training guide: Midjourney, specifically for beauty content.
Each chapter takes one shot type a beauty team actually needs, such as a dropper close-up, a product in hand, a bathroom setting, or a texture close-up. It shows a recommended base prompt, the target image, and three steps: analyze a reference with /describe, fix flawed areas with vary(region), then push realism with remix mode and --style raw. Every step keeps the full prompt and the reason for the change, so a teammate can repeat it without me in the room.
Step 1
Read the reference
Run /describe on a reference image and start from the prompt it returns.
Step 2
Fix only what's wrong
Select the flawed area with vary(region) and regenerate just that part.
Step 3
Push the realism
Add --style raw in remix mode, then repeat with the closest result.
Date
2024.04
Tooll
Midjourney v6
08 — Collaboration
Support additional shoots and 3D work.
Alongside my AI content responsibilities, I volunteered to support a Staaack shoot and proposed and completed additional 3D work. I reviewed deadlines and priorities to manage the work alongside my existing tasks.
STAAACK
A reusable 3D asset kit for LG H&H STAAACK Magazine Issue 23, built to sit naturally inside live cover photography.
LG생활건강 STAAACK 매거진 23호를 위한 재사용 가능한 3D 에셋 키트. 실사 커버 사진 안에 자연스럽게 어우러지도록 제작했습니다.
Period · 기간
April 2024 · 2024년 4월
Role · 역할
3D · Art Direction · Compositing · 3D · 아트 디렉션 · 합성
Tools · 툴
Cinema 4D · Redshift · Photoshop
A new magazine needed a cover language.
The request was 3D assets for the cover. The harder part: STAAACK was still defining its visual language, and the assets had to sit on live photography without looking pasted on. So the mood had to be read from the actual shoot, not a reference board.
Step · 02
Switch renderer,
keep the deadline.
The studio pipeline used Redshift, while my previous workflow was Octane. I switched renderers inside the project and rebuilt the shader logic quickly enough to keep production moving.
Tool · Redshift, first production use
Step · 03
Build assets
for a live plate.
Typography, liquid, character, bubble, and dollar-stack assets were built to merge with one photographic plate. Each asset was lit, graded, and framed as if it had already been on set.
Output · six or more assets, one composite cover
Step · 03 — Asset board · 에셋 보드
Three type treatments, built before the composite
합성 전에 먼저 만든 세 가지 타이포그래피 에셋
01 · LIQUID TYPE · 리퀴드 타이포
Four S forms, from scattered droplets to one continuous pour.
흩어진 물방울부터 하나로 이어진 흐름까지, 네 가지 S 형태.
02 · INFLATED TYPE · 풍선 타이포
One phrase in three colorways and angles.
같은 문구를 세 가지 컬러와 각도로.
03 · BLOCK TYPE · 블록 타이포
Letter blocks in three arrangements: stacked, staggered, rising.
글자 블록을 세 가지 배열로: 겹침, 엇갈림, 상승.
09 — Outcome
Outcomes and learning
결과와 배운 점
Leave the team with both content and a documented process.
Learning
I proposed and applied a workflow connecting product assets, scene setup, AI generation, and output selection. I documented it in walkthrough videos and training materials for internal use.
The experience taught me to decide what could change and what needed to stay accurate before using AI.
I learned to apply my 3D skills to problems in AI production as well as to the visuals themselves. Documenting the work also helped me explain the process clearly to colleagues.
Next Project · 다음 프로젝트
Bubblify
→
Copyright © Sydney Hong 2026
Try
Framer
for Free!