Overview
Research addresses the problem of source attribution for generative 3D models, focusing on identifying the specific model responsible for creating a given 3D asset. This attribution is critical for applications in gaming, robotics, and immersive creation.
Research Context
The challenge of source attribution for generative 3D assets is characterized by two core issues: dispersed attribution signals and realistic deployment constraints. Dispersed attribution signals imply that identifying characteristics, or "3D fingerprints," are distributed across various data modalities, including multi-view, geometric, and frequency-domain cues. Realistic deployment constraints involve scenarios with scarce labels, degraded prompts, and mixtures of real and synthetic assets, which can undermine the reliability of attribution efforts.
To systematically study this problem, a new passive source attribution benchmark was constructed. This benchmark is designed for modern generated assets and covers 22 representative 3D generators. It includes evaluation under standard, few-shot, and realistic deployment protocols.
Approach
The research first centered on characterizing the types of stable fingerprints left by generative 3D models. Based on this analysis, generative 3D models were found to leave two primary types of stable fingerprints:
- Cross-view inconsistency: Variations or discrepancies in how an asset appears from different viewpoints.
- Structural artifacts: These are reflected in geometric statistics and frequency-domain cues, indicating characteristic imperfections or patterns introduced by the generative process.
To capture these dispersed signals comprehensively, the researchers proposed a hierarchical multi-view multi-modal Transformer. This architecture is designed to:
- Fuse appearance, geometric, and frequency-domain features within each individual view of a 3D asset.
- Model global relationships across multiple views, integrating information from different perspectives.
The benchmark and the proposed method were evaluated under various data availability conditions, including full supervision and scenarios with limited training data (few-shot learning, specifically 1% of the training data).
Findings
The experimental results demonstrated strong performance in attributing the source of generative 3D models. Key findings include:
- Generative 3D models leave stable and attributable fingerprints, identifiable as cross-view inconsistency and structural artifacts in geometric statistics and frequency-domain cues.
- The proposed hierarchical multi-view multi-modal Transformer achieved an accuracy of 97.22% under full supervision on the constructed benchmark.
- Under conditions of limited training data, specifically with only 1% of the training data (corresponding to fewer than five samples per generator), the method achieved an accuracy of 77.17%.
These findings indicate that modern 3D generators produce consistent, identifiable characteristics, allowing for their provenance to be traced.
Why This Matters
Source attribution for generative 3D models is critical for applications in gaming, robotics, and immersive creation. Establishing the ability to identify the generative model that created a 3D asset contributes to trustworthy 3D content provenance. This systematic study and proposed methodology establish a new benchmark and a methodological foundation for addressing this problem.