Scholarlink

<p>Face aging synthesis aims to generate plausible future facial appearances while preserving the subject’s original identity, with applications ranging from missing-person searches to digital-twin construction. However, achieving a balance between maintaining identity and introducing realistic age-related geometric remains a fundamental challenge. Existing methods, including GAN-based and recent diffusion-based approaches, often rely on single-modality guidance, leading to trade-offs between aging accuracy and identity fidelity. To address this, we propose multi-modal diffused face aging (MDFA), a novel adapter-based framework that synergistically integrates dense identity features with sparse geometric priors within a frozen latent diffusion backbone. Our architecture introduces a multi-level cross-attention (MLCA) module to harmonize these heterogeneous modalities across multiple scales, alongside a latent adapter (LA) that hierarchically injects the fused guidance into the diffusion U-Net. This design enables precise control over structural aging transformations while preserving fine-grained identity cues. Extensive experiments on the FFHQ dataset demonstrate that MDFA achieves state-of-the-art performance, with an FID of 31.86 and an identity similarity score of 0.4972, significantly outperforming existing methods in both quantitative metrics and human perceptual evaluations. The proposed approach advances the field of facial age progression by offering a principled, multi-modal fusion strategy that effectively reconciles identity preservation with realistic aging synthesis, paving the way for more reliable and controllable digital human applications. The code is available at: <a href="https://github.com/SouthDridge/MDFA-Multi-modal-Diffused-Face-Aging">https://github.com/SouthDridge/MDFA-Multi-modal-Diffused-Face-Aging</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Synergistic multi-modal fusion for identity-preserving face aging via latent diffusion

  • 期刊论文

  • 发表日期：

    2026年4月19日

查看全文
  • Chenheng Deng,
  • Yiran Peng,
  • Yiyao Huang,
  • Ruijia Lei,
  • Kin Tak U,
  • Wenxiao Wang

摘要

Face aging synthesis aims to generate plausible future facial appearances while preserving the subject’s original identity, with applications ranging from missing-person searches to digital-twin construction. However, achieving a balance between maintaining identity and introducing realistic age-related geometric remains a fundamental challenge. Existing methods, including GAN-based and recent diffusion-based approaches, often rely on single-modality guidance, leading to trade-offs between aging accuracy and identity fidelity. To address this, we propose multi-modal diffused face aging (MDFA), a novel adapter-based framework that synergistically integrates dense identity features with sparse geometric priors within a frozen latent diffusion backbone. Our architecture introduces a multi-level cross-attention (MLCA) module to harmonize these heterogeneous modalities across multiple scales, alongside a latent adapter (LA) that hierarchically injects the fused guidance into the diffusion U-Net. This design enables precise control over structural aging transformations while preserving fine-grained identity cues. Extensive experiments on the FFHQ dataset demonstrate that MDFA achieves state-of-the-art performance, with an FID of 31.86 and an identity similarity score of 0.4972, significantly outperforming existing methods in both quantitative metrics and human perceptual evaluations. The proposed approach advances the field of facial age progression by offering a principled, multi-modal fusion strategy that effectively reconciles identity preservation with realistic aging synthesis, paving the way for more reliable and controllable digital human applications. The code is available at: https://github.com/SouthDridge/MDFA-Multi-modal-Diffused-Face-Aging