Design of a Privacy-Preserving Local Virtual Try-On System Using Diffusion-Based Generative Models
Author
Wollner Infante Papa, Vittorio
Term
4. semester
Publication year
2026
Submitted on
2026-05-31
Pages
69
Abstract
Online shopping sees high return rates because buyers are unsure how clothes will fit. Virtual try-on (VTO) tools let people preview garments on their own body before buying, which could reduce returns. Yet the best current VTO systems run in the cloud or use third-party APIs. That conflicts with GDPR, because customer photos are biometric data with special protections. This thesis designs and tests a different approach: a fully self-contained VTO system that runs only on local hardware, with no external services, to protect privacy. Following the Design Science Research paradigm, the work builds a virtual try-on pipeline using FLUX.2-klein-9B, a 9-billion-parameter Diffusion Transformer (an image-generation model), adapted with Low-Rank Adaptation (LoRA), and deployed on a single NVIDIA A40 GPU. The pipeline includes a dedicated preprocessing and user-normalization stage to clean and standardize real customer photos, addressing a common weak point in earlier VTO research. The system was compared with commercial baselines (GPT-4o from OpenAI and Gemini from Google) in a blind perceptual study. Twelve participants judged three aspects: output realism, garment fidelity (how well the clothing matches the target item), and identity retention (does it still look like the same person). A nonparametric Sign Test found no statistically significant preference for the commercial systems on any measure. On garment fidelity, the local system was numerically preferred, but the difference was not significant given the small sample. These results show that privacy-preserving, locally deployed diffusion-based VTO can achieve perceived quality comparable to widely available multimodal AI services, while avoiding the regulatory risk of sending biometric data to third parties. The study is a pilot; larger user studies and evaluation in live retail settings are proposed next steps.
Webbutikker oplever høje returneringsrater, fordi kunder er usikre på, om tøjet passer. Virtuel prøvning (VTO) lader folk se tøjet på deres egen krop før køb og kan dermed reducere returneringer. De bedste VTO-løsninger i dag kører dog i skyen eller via tredjeparts-API'er, hvilket kolliderer med GDPR, da kundebilleder er biometriske data med særlige beskyttelser. Denne afhandling udvikler og afprøver et alternativ: et fuldt selvstændigt VTO-system, der udelukkende kører på lokal hardware uden eksterne tjenester for at beskytte privatliv. Inden for Design Science Research-paradigmet bygges en virtuel prøvningspipeline med FLUX.2-klein-9B, en Diffusion Transformer til billedgenerering med 9 milliarder parametre, tilpasset med Low-Rank Adaptation (LoRA) og kørt på et enkelt NVIDIA A40-GPU-kort. Pipeline'en omfatter dedikeret forbehandling og bruger-normalisering for at rense og standardisere virkelige kundebilleder og dermed adressere en udbredt svaghed i tidligere VTO-arbejder. Artefaktet blev sammenlignet med kommercielle baseline-systemer (GPT-4o fra OpenAI og Gemini fra Google) i en blind perceptuel undersøgelse. Tolv deltagere vurderede tre forhold: output-realisme, tøjets overensstemmelse (garment fidelity) og bevarelse af identitet (ligner det den samme person). En ikke-parametrisk Sign Test fandt ingen statistisk signifikant præference for de kommercielle systemer på nogen dimension. På tøjets overensstemmelse blev den lokale løsning foretrukket numerisk, men forskellen var ikke signifikant givet den lille stikprøve. Resultaterne peger på, at privatlivsbevarende, lokalt deployerede, diffusionsbaserede VTO-systemer kan opnå en oplevet kvalitet på niveau med udbredte multimodale AI-tjenester, samtidig med at man undgår regulatorisk risiko ved at sende biometriske data til tredjepart. Undersøgelsen er en pilot; næste skridt er større brugerundersøgelser og evaluering i drift i en detailkontekst.
[This apstract has been rewritten with the help of AI based on the project's original abstract]