ArtArchitect: Adaptive Depth Correction and Outpainting for Realistic World Generation

We address the challenge of creating seamless, visually coherent 3D scenes from text descriptions and images, a task that is frequently hindered by sudden transitions, boundary artifacts, and inconsistent depth representations in current techniques. These issues not only disrupt the realism but also worsen as the scene scales up, resulting in poor quality for large-scale scene generation. To overcome these challenges, we present ArtArchitect, an innovative framework that introduces two main advancements. First, an enhanced Outpaint Modulator utilizes channel-aware noise thresholding to ensure smooth transitions, reduce boundary artifacts, and maintain fine details. Second, an Adaptive Depth Modulator combines adaptive window-based depth filling with edge-preserving smoothing to generate consistent depth maps and preserve geometric stability across different views. Our method effectively addresses common artifacts, enabling the creation of high-quality, large-scale 3D environments with strong spatial and visual continuity. Both qualitative and quantitative evaluations show that ArtArchitect significantly outperforms existing methods, setting a new benchmark for stable, immersive, and visually impressive 3D scene generation (Project page).