Research on Virtual Clothing Try-On Using Stable Diffusion and LoRA

Background: AI innovations have made tools like Stable Diffusion popular for image generation. This work explores using Stable Diffusion for virtual clothing try‑on.

Two main approaches are compared: traditional VTON pipelines (geometric matching + try‑on module) and AIGC‑based methods that leverage Stable Diffusion with extensions such as LoRA, ControlNet, Inpainting, and Segment Anything.

Methodology: LoRA models are fine‑tuned on ~10 images per garment after background removal and resizing to 512×512. ControlNet provides pose and edge conditioning; Inpainting allows masked region regeneration; Segment Anything assists in mask creation. The workflow is implemented on a local Stable Diffusion Web UI (AUTOMATIC1111) running on an RTX 3060 12 GB GPU, with plugins installed via the extensions folder.

Training: Approximately 30 minutes per LoRA model (≈37 MB) for 50 epochs on a single GPU. Tag generation is performed with the Tagger plugin, and tags are edited using Dataset Tag Editor to improve data quality.

Results: Single‑item try‑on (t‑shirts, jackets, shoes, etc.) shows high fidelity in color and texture. Seasonal outfit models (spring‑autumn, summer, winter) also produce reasonable compositions. Overall, the AIGC pipeline achieves comparable or better visual quality than conventional VTON, especially for upper‑body garments.

Conclusion: By locally deploying Stable Diffusion and augmenting it with LoRA and ControlNet, a practical virtual try‑on demo is realized, demonstrating the potential of AI‑generated fashion applications.