کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

PC-MCLD: Pose-Constrained and Multi-focal Conditioned Latent Diffusion for Person Image Synthesis

نویسندگان :

Hanieh Fazli¹ Reza Azmi²

1- دانشگاه الزهرا(س) 2- دانشگاه الزهرا(س)

کلمات کلیدی :

pose-guided person image synthesis،latent diffusion model،texture consistency،adaptive feature fusion،fashion image generation

چکیده :

Pose-guided person image synthesis (PGPIS) aims to generate a person in a target pose while preserving identity and garment details, yet large pose variations often cause texture misalignment and loss of facial fidelity in existing diffusion models. We propose PC-MCLD, a latent diffusion framework that introduces (i) a pose-aware texture transfer constraint ensuring anatomically consistent correspondence between source and target regions, and (ii) an adaptive weighting mechanism that balances global appearance, garment texture, and facial identity cues during generation. Experiments on the DeepFashion In-Shop benchmark show clear improvements over a reproduced MCLD baseline. At 176×256, PC-MCLD reduces FID by 1.39% and LPIPS by 8.24%; at 352×512, the gains increase to 2.53% in FID and 19.48% in LPIPS. These results demonstrate that PC-MCLD enhances both perceptual quality and structural fidelity under challenging pose changes.