ABSTRACT Generative artificial intelligence is increasingly used in creative industries, yet its application in fashion design often remains limited to visual generation rather than structured design support. This study proposes a human‐in‐the‐loop controllable diffusion framework for virtual fashion prototyping. The pipeline integrates user‐refined garment masks, pose‐guided structural conditioning, sketch‐to‐render edge constraints, and lightweight LoRA style adaptation, enabling localized editing while preserving global body–garment coherence. Unlike prompt‐only image generation, conventional virtual try‐on systems, or downstream 3D production tools, the framework functions as a pre‐production visual ideation layer for designer‐led iterative editing. The study evaluates the framework through quantitative comparison with directly comparable GAN‐ and diffusion‐based baselines on DeepFashion2, using FID, LPIPS, SSIM, and Human Score; ablation analysis of individual control components; and a within‐subject user study design involving professional fashion designers across representative design tasks. A pilot evaluation with six experienced designers is also reported. Results show that the proposed framework outperforms directly comparable baselines across all reported metrics. Pilot findings further indicate higher perceived control, lower workload, and stronger usability than prompt‐only and inpainting baselines. The findings suggest that controllable diffusion can support professional fashion ideation as an interactive prototyping system rather than merely an automated image generator.
Guopeng Zhang (2026) studied this question.