RoamScene3D: Immersive Text-to-3D Scene Generation via Adaptive Object-aware Roaming

TL;DR

This work proposes RoamScene3D, a novel framework that bridges the gap between semantic guidance and spatial generation, and significantly outperforms state-of-the-art approaches in producing consistent and photorealistic scenes.

Abstract

Generating immersive 3D scenes from texts is a core task in computer vision, crucial for applications in virtual reality and game development. Despite the promise of leveraging 2D diffusion priors, existing methods suffer from spatial blindness and rely on predefined trajectories that fail to exploit the inner relationships among salient objects. Consequently, these approaches are unable to comprehend the semantic layout, preventing them from exploring the scene adaptively to infer occluded content. Moreover, current inpainting models operate in 2D image space, struggling to plausibly fill holes caused by camera motion. To address these limitations, we propose RoamScene3D, a novel framework that bridges the gap between semantic guidance and spatial generation. Our method reasons about the semantic relations among objects and produces consistent and photorealistic scenes. Specifically, we employ a vision-language model (VLM) to construct a scene graph that encodes object relations, guiding the camera to perceive salient object boundaries and plan an adaptive roaming trajectory. Furthermore, to mitigate the limitations of static 2D priors, we introduce a Motion-Injected Inpainting model that is fine-tuned on a synthetic panoramic dataset integrating authentic camera trajectories, making it adaptive to camera motion. Extensive experiments demonstrate that with semantic reasoning and geometric constraints, our method significantly outperforms state-of-the-art approaches in producing consistent and photorealistic scenes. Our code is available at https://github.com/JS-CHU/RoamScene3D.

Figures (9)

• Figure 1: The visual results of the proposed method. Given a text description, our method is capable of generating arbitrary indoor and outdoor scenes. The synthesized environments are photorealistic and immersive, enabling navigable walkthroughs. The figure illustrates a trajectory transitioning from one viewpoint to another within the scene.
• Figure 2: Overview of the RoamScene3D framework. Given a text prompt, our method first initializes the scene by generating an RGBD panorama and unprojecting it into a coarse 3D representation. To ensure semantic fidelity, we utilize a VLM to construct a Semantic Scene Graph, which encodes object relationships and guides to adaptively generate a closed, object-aware camera trajectory. Subsequently, we employ a Motion-injected Panorama RGBD Inpainting model to synthesize consistent novel views conditioned on camera movement, effectively handling disocclusions. Finally, the sequence of rendered panoramas is integrated via 3D Gaussian Splatting optimization to produce a photorealistic and spatially coherent immersive 3D scene.
• Figure 3: The visualization of the generated adaptive trajectory. It is plotted in a 2D plane for simplification. The outermost gray rectangle denotes the scene boundary, while the colored rectangles represent objects within the scene. The figure illustrates the evolution from the initial trajectory (gray dashed line) to the final trajectory (solid black line) following the applied perturbation (pink).
• Figure 4: The architecture of the proposed motion-injected panorama inpainting model. We design a camera motion encoder to inject movement into the UNet. To train the interaction between static and dynamic features, we use LoRA adapters to fine-tune all the self-attention and cross-attention layers.
• Figure 5: The visual examples of our inpainting and SR enhancement operation. The novel rendered views undergo inpainting and super-resolution to generate the final reference images. In the dashed box, we compare the inpainting results with and without motion injection.