Scholarlink

We introduce FLAME_GAN, a novel framework for high-fidelity and controllable 3D face reconstruction from a single 2D image, designed to support identity-aware retrieval and similarity-based applications. Our approach combines the strengths of Generative Adversarial Networks (GANs), the FLAME parametric 3D face model, and 3D Morphable Model (3DMM) priors to achieve robust and interpretable facial synthesis. By extending the DiscoFaceGAN architecture, we disentangle identity, expression, and pose in the latent space, enabling explicit and fine-grained control. The FLAME model introduces structural awareness, while 3DMM priors guide geometric consistency through tailored loss functions. Our loss formulation includes adversarial, identity, reconstruction, geometric consistency, and semantic parsing losses. Experiments on FFHQ and CelebA benchmarks demonstrate that FLAME_GAN outperforms state-of-the-art methods in both visual fidelity and structural accuracy. Furthermore, we evaluate identity similarity under pose and expression variations, confirming the suitability of the model for face re-identification and similarity search tasks. These results position FLAME GAN as a competitive solution in both facial reconstruction and multimedia retrieval contexts.

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

FLAME_GAN: High-Fidelity 3D Face Reconstruction via Disentangled Representations and Explicit Geometric Control

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Grina Wiem,
  • Douik Ali

摘要

We introduce FLAME_GAN, a novel framework for high-fidelity and controllable 3D face reconstruction from a single 2D image, designed to support identity-aware retrieval and similarity-based applications. Our approach combines the strengths of Generative Adversarial Networks (GANs), the FLAME parametric 3D face model, and 3D Morphable Model (3DMM) priors to achieve robust and interpretable facial synthesis. By extending the DiscoFaceGAN architecture, we disentangle identity, expression, and pose in the latent space, enabling explicit and fine-grained control. The FLAME model introduces structural awareness, while 3DMM priors guide geometric consistency through tailored loss functions. Our loss formulation includes adversarial, identity, reconstruction, geometric consistency, and semantic parsing losses. Experiments on FFHQ and CelebA benchmarks demonstrate that FLAME_GAN outperforms state-of-the-art methods in both visual fidelity and structural accuracy. Furthermore, we evaluate identity similarity under pose and expression variations, confirming the suitability of the model for face re-identification and similarity search tasks. These results position FLAME GAN as a competitive solution in both facial reconstruction and multimedia retrieval contexts.