Problem
Users wanted personalized artwork that retained recognizable identity across different styles when mainstream image tools could not yet do this reliably.
04 / CASE STUDY
Generative imaging · Personalized diffusion
Mokki Art
An identity-preserving image generation system shipped before high-quality personalized generation became a commodity feature.
RoleSenior generative AI engineer
StatusDecommissioned product
Timeline2021 — 2023
01 / CONTEXTThe problem
The problem
behind the product.
Approach
The system used a small set of user-provided facial images to create a personalized generation workflow, with attention to preprocessing, identity consistency and GPU inference.
02 / SYSTEMArchitecture,
Architecture,
not magic.
01Image intake→
02Preprocessing→
03Personalization→
04Diffusion inference→
05Quality selection
Constraints
- Only 6–12 source images
- Early diffusion tooling
- GPU inference cost
- Identity consistency
My ownership
- Designed the personalized generation pipeline
- Implemented image preprocessing and model workflows
- Integrated GPU inference into a customer product
- Improved consistency across styles and outputs
03 / EVIDENCEWhat can be
What can be
defended.
Shipped customer-facing product
Personalization from 6–12 photos
Identity-preserving workflow
Delivered before the capability became widely commoditized
What I would improve now.
The product was ultimately decommissioned as frontier tools made personalized generation broadly accessible. The architecture remains evidence of early delivery under immature tooling.
PyTorchStable DiffusionComputer visionGPU inferenceControlNet