近年來，生成式模型在影像生成與編輯領域展現出優異的成果，其中影像重光照技術能夠模擬相同場景在不同光照條件下的外觀變化，具有虛擬場景生成、影像編輯與電腦視覺等多項應用潛力。然而，現有方法在光照控制能力、場景結構保持以及生成影像細節保留等方面仍面臨挑戰。因此，本研究以生成式模型為基礎，探討室內場景影像重光照方法，並分析不同模型設計對影像品質與光照控制效果的影響。 本研究首先回顧影像重光照相關技術，包括生成對抗網路、擴散模型、潛在空間控制以及條件式生成模型，分析其在光照建模與可控生成上的優勢與限制。接著，以 LSUN 資料集作為主要訓練資料來源，並結合 MIT Multi-Illumination Dataset提供之不同光照條件影像，建立室內場景重光照實驗環境，以學習場景內容與光照變化之間的關係。 在模型設計方面，本研究提出以 ConvNeXt-V2 結合變分自編碼器（Variational Au-toencoder， VAE）作為編碼器，搭配預先訓練且參數固定的 StyLitGAN 生成器所組成之 Variational-StyLitGAN 架構。模型將輸入影像映射至潛在空間後，透過光照控制向量調整生成結果，以實現不同光照條件下的影像重建與重光照，同時利用 Mean Squared Error（MSE）、Learned Perceptual Image Patch Similarity（LPIPS）以及 Kullback-Leibler（KL） Divergence等損失函數共同最佳化，以兼顧重建準確性、感知品質與潛在空間穩定性。 實驗結果顯示，所提出的方法能在維持場景幾何結構與物件配置一致性的前提下，生成具有不同光照效果的室內場景影像，並於影像細節保留與視覺品質方面展現良好表現，驗證了結合 ConvNeXt-VAE 與 StyLitGAN 進行可控式影像重光照之可行性。 關鍵字：影像重光照、生成式模型、生成對抗網路、變分自編碼器、潛在空間控制、StyLitGAN、ConvNeXt-V2;In recent years， generative models have achieved remarkable success in image generation and editing tasks. Among these applications， image relighting aims to synthesize realistic images of the same scene under different illumination conditions and has attracted increas-ing attention in computer vision research. However， existing methods still face challenges in lighting controllability， structural consistency， and preservation of fine visual details. There-fore， this study investigates a controllable image relighting framework based on generative models and analyzes the influence of different architectural designs on image quality and illumination control. This research first reviews representative image relighting techniques， including Generative Adversarial Networks (GANs)， Diffusion Models， latent space manipulation， and condition-guided generation methods， discussing their advantages and limitations in illumination modeling and controllable image synthesis. For model training， the LSUN dataset is adopted as the primary source of indoor scene images， while the MIT Multi-Illumination Dataset is utilized to provide reference images captured under different lighting conditions for relight-ing-related analysis and evaluation. The proposed framework is built upon a Variational-StyLitGAN architecture that combines a ConvNeXt-V2 -based Variational Autoencoder (VAE) encoder with a pretrained and fro-zen StyLitGAN generator. The encoder maps input images into a latent representation， and controllable relighting is achieved by manipulating lighting condition vectors in the genera-tor′s latent space. To improve reconstruction fidelity， perceptual quality， and latent-space regularization， the model is jointly optimized using Mean Squared Error (MSE) loss， Learned Perceptual Image Patch Similarity (LPIPS) loss， and Kullback-Leibler (KL) diver-gence. Experimental results demonstrate that the proposed framework can generate indoor scene images with different illumination conditions while preserving scene geometry and object arrangements. Furthermore， incorporating perceptual loss improves edge sharpness and tex-ture preservation， resulting in more visually realistic reconstruction and relighting perfor-mance. These findings verify the feasibility of combining ConvNeXt-V2， VAE， and Sty-LitGAN for controllable image relighting and provide insights into balancing reconstruction accuracy and perceptual image quality through latent-space learning. Keywords: Image Relighting， Generative Models， Generative Adversarial Networks， Varia-tional Autoencoder， Latent Space Manipulation， StyLitGAN， ConvNeXt-V2
