BlockFusion: Expandable 3D Scene Generation using Latent Tri-plane Extrapolation (2401.17053v4)

Abstract: We present BlockFusion, a diffusion-based model that generates 3D scenes as unit blocks and seamlessly incorporates new blocks to extend the scene. BlockFusion is trained using datasets of 3D blocks that are randomly cropped from complete 3D scene meshes. Through per-block fitting, all training blocks are converted into the hybrid neural fields: with a tri-plane containing the geometry features, followed by a Multi-layer Perceptron (MLP) for decoding the signed distance values. A variational auto-encoder is employed to compress the tri-planes into the latent tri-plane space, on which the denoising diffusion process is performed. Diffusion applied to the latent representations allows for high-quality and diverse 3D scene generation. To expand a scene during generation, one needs only to append empty blocks to overlap with the current scene and extrapolate existing latent tri-planes to populate new blocks. The extrapolation is done by conditioning the generation process with the feature samples from the overlapping tri-planes during the denoising iterations. Latent tri-plane extrapolation produces semantically and geometrically meaningful transitions that harmoniously blend with the existing scene. A 2D layout conditioning mechanism is used to control the placement and arrangement of scene elements. Experimental results indicate that BlockFusion is capable of generating diverse, geometrically consistent and unbounded large 3D scenes with unprecedented high-quality shapes in both indoor and outdoor scenarios.

• The paper presents a latent tri-plane diffusion model that generates high-quality 3D scenes by representing structures as expandable unit blocks.
• It employs an extrapolation mechanism that integrates new blocks using overlapping latent features to maintain geometric consistency across scenes.
• A 2D layout conditioning method enables precise control over scene element placement, enhancing applications in gaming, AR/VR, and filmmaking.

Introduction

3D scene generation has become a critical area of research, fueled by its applications in industries such as gaming, filmmaking, and AR/VR. While 2D diffusion models have achieved remarkable successes in image synthesis, translating these advancements to the 3D domain presents unique challenges. Traditional methods have predominantly focused on generating 3D content within a fixed spatial extent—leaving the task of creating expandable 3D scenes relatively unexplored.

BlockFusion: Innovation in 3D Scene Generation

The paper introduces BlockFusion, an innovative model leveraging a tri-plane based architecture to generate and expand 3D scenes in a coherent and systematic manner. This method operates by generating scenes as unit blocks and enabling seamless incorporation of new blocks to extend the scene. Uniquely, BlockFusion is trained on datasets of randomly cropped 3D blocks from complete scene meshes, extracting high-level features into a hybrid neural field structure before encoding these into a more compact latent space for diffusion.

Key Contributions

The method proposed brings three significant advancements to the table:

1. A latent tri-plane diffusion model that generalizes for high-quality 3D shape generation at the scene level.
2. An extrapolation mechanism that allows for harmonious expansion of scenes by conditioning the generative process on features from existing blocks' latent representations.
3. A 2D layout conditioning mechanism that grants users control over the placement and configuration of scene elements.

Experimental results underscore BlockFusion's capability of generating geometrically consistent and diverse large 3D scenes, outperforming methods that either generate fixed-size content or utilize less direct approaches for extendable scenes.

Methodology

BlockFusion's approach entails several steps, starting from the transformation of training 3D blocks to raw tri-planes, compression into a latent tri-plane space, and subsequently training a denoising diffusion probabilistic model (DDPM). To extend a scene, BlockFusion extrapolates these latent tri-planes to new blocks, syncing the generation process with overlapping features. The method also incorporates a straightforward 2D layout input to guide the arrangement of objects within the scene.

Performance

The paper presents robust quantitative results, revealing BlockFusion's strong generative performance. The approach demonstrates the ability to generate scenes with consistent geometry and novel layouts that weren’t present in the training set. The method outperforms previous work in creating large-scale scenes and shows potential as an industry-level content creation tool. However, it's worth noting the method's limitations in generating finer geometric details and the challenge of producing consistent textures for expansive scenes—areas identified for future enhancement.

BlockFusion embodies a significant step towards the automated creation of large-scale 3D content, with potential applications ranging from video game design to virtual exploration. Its expandable nature breaks free from the confines of predetermining world boundaries, providing a sophisticated toolset for the ongoing development in the field of generative 3D modeling.