TIMI: Training-Free Image-to-3D Multi-Instance Generation with Spatial Fidelity

TL;DR

TIMI is proposed, a novel Training-free framework for Image-to-3D Multi-Instance generation that achieves high spatial fidelity and yields better performance in terms of both global layout and distinct local instances compared to existing multi-instance methods, without requiring additional training and with faster inference speed.

Abstract

Precise spatial fidelity in Image-to-3D multi-instance generation is critical for downstream real-world applications. Recent work attempts to address this by fine-tuning pre-trained Image-to-3D (I23D) models on multi-instance datasets, which incurs substantial training overhead and struggles to guarantee spatial fidelity. In fact, we observe that pre-trained I23D models already possess meaningful spatial priors, which remain underutilized as evidenced by instance entanglement issues. Motivated by this, we propose TIMI, a novel Training-free framework for Image-to-3D Multi-Instance generation that achieves high spatial fidelity. Specifically, we first introduce an Instance-aware Separation Guidance (ISG) module, which facilitates instance disentanglement during the early denoising stage. Next, to stabilize the guidance introduced by ISG, we devise a Spatial-stabilized Geometry-adaptive Update (SGU) module that promotes the preservation of the geometric characteristics of instances while maintaining their relative relationships. Extensive experiments demonstrate that our method yields better performance in terms of both global layout and distinct local instances compared to existing multi-instance methods, without requiring additional training and with faster inference speed.

Figures (10)

• Figure 1: Comparison of our training-free multi-instance 3D generation method with existing approaches. TIMI generates multiple 3D instances from a single image by training-free guiding pre-trained 3D diffusion models, achieving precise global layout and distinct local instances, without requiring additional fine-tuning.
• Figure 2: Overview of the TIMI framework. Given a single image and instance masks, TIMI guides a frozen pre-trained Image-to-3D diffusion model to generate multi-instance 3D outputs without additional training. (a) Instance-aware Separation Guidance applies instance-level constraints to early cross-attention layers to promote instance separation. (b) Spatial-stabilized Geometry-adaptive Update stabilizes inference-time guidance via geometry-adaptive gradient modulation to preserve overall spatial structure.
• Figure 3: Qualitative Comparisons on Synthetic Data. Existing methods often produce inaccurate global layouts or fail to separate local instances. Our method preserves both global and local spatial fidelity, yielding well-disentangled instances.
• Figure 4: Qualitative comparisons on real and stylized images.
• Figure 5: Qualitative ablation study on the effectiveness of proposed components in TIMI. The results demonstrate that each module progressively enhances the layout alignment and instance distinctiveness, with the full method achieving the best spatial fidelity.