Our project focuses on refining the process of 3D modeling from multi-view images, aiming to enhance the final model’s fidelity and texture details while maintaining computational efficiency. The core of our methodology involves our own refiner model that preprocesses images to significantly upscales the clarity, detail, textures and lightning, which will be very much benefitial for mapping the details from the images for 3D mesh generation. By implementing techniques like VAE and Unets, which optimizes texture and geometric details from these source images, we ensure that the generated 3D models are high in quality and also rich in the textures based on the original source images and even consistent enough to generate a 3D mesh using NGP techniques. Results from our experiments demonstrate that our approach improves the visual accuracy of the 3D models with remarkably low GPU resources. This advancement underscores the potential for applying our methods in real-time 3D modeling applications, opening new way for both practical and innovative uses.