REPARO: Compositional 3D Assets Generation with Differentiable 3D Layout Alignment

TL;DR

REPARO tackles multi-object 3D generation from a single image by decomposing the scene into individually reconstructed objects and globally aligning their layout with differentiable rendering. It introduces an optimal transport-based long-range appearance loss and a high-level semantic loss to enforce global correspondences in RGB, depth, and semantic features. The two-stage pipeline yields high-fidelity per-object assets and coherent scene layouts, outperforming baselines on multi-object datasets and in-the-wild images, with a user study favoring REPARO. The approach reduces post-processing and enables instance-level meshes aligned to input images, though it can hallucinate under severe occlusion or semantic ambiguities.

Abstract

Traditional image-to-3D models often struggle with scenes containing multiple objects due to biases and occlusion complexities. To address this challenge, we present REPARO, a novel approach for compositional 3D asset generation from single images. REPARO employs a two-step process: first, it extracts individual objects from the scene and reconstructs their 3D meshes using off-the-shelf image-to-3D models; then, it optimizes the layout of these meshes through differentiable rendering techniques, ensuring coherent scene composition. By integrating optimal transport-based long-range appearance loss term and high-level semantic loss term in the differentiable rendering, REPARO can effectively recover the layout of 3D assets. The proposed method can significantly enhance object independence, detail accuracy, and overall scene coherence. Extensive evaluation of multi-object scenes demonstrates that our REPARO offers a comprehensive approach to address the complexities of multi-object 3D scene generation from single images.

Figures (7)

• Figure 1: Qualitative comparison of generated 3D assets of the single object and multiple objects. (b) and (c) are generated 3D assets of the single object assets using DreamGaussian tang2023dreamgaussian; (d) is a 3D asset of multiple objects generated by DreamGaussian tang2023dreamgaussian; (e) is our result.
• Figure 2: The diagram of the proposed REPARO. (a) is the pipeline to reconstruct the 3D asset of each object in the reference image. $B$ and $O$ denote bounding boxes and occlusion information of each object, respectively. If an object is occluded, the preprocessing module will complement it using the inpainting model. (b) is the process of layout alignment based on differentiable rendering. The parameters of reconstructed meshes are optimized by gradient descent. The loss function $L$ (Eq. \ref{['eq:total_loss']}) consists of the long-range appearance loss $L_a$ and the high-level semantic loss $L_s$.
• Figure 3: GPU memory usage and approximate elapsed time for optimization of compositional assets layout under different parameter settings.
• Figure 4: Qualitative comparison with different image-to-3D generation models. Given an input image, previous methods produce inaccurate textures and geometry with noticeable artifacts. Our method generates high-quality, high-fidelity compositional assets with the correct spatial layout.
• Figure 5: Qualitative comparison for different options of optimization parameters $\theta$.