RecDreamer: Consistent Text-to-3D Generation via Uniform Score Distillation

Prevailing text-to-3D generation methods based on score distillation often suffer from significant geometric inconsistencies, manifesting as repeated patterns across different poses of 3D assets—an issue commonly referred to as the Multi-Janus problem. Although recent work has attempted to address this by improving control over pose prompts or adjusting the approximation distribution, the underlying prior distribution remains biased toward a canonical pose, leading to skewed guidance.

To overcome this challenge of inconsistency due to an imbalanced prior distribution, we have investigated techniques for modifying a prescribed distribution, enabling reconstruction of its density to ensure compliance with specific marginal constraints. By modifying the original data distribution through the proposed auxiliary function, we ensure that the marginal distribution of the pose adheres to a uniform distribution, thereby eliminating biases from prior knowledge. We integrate the rectified data distribution into existing score distillation algorithms, and term the process uniform score distillation.

To efficiently compute the posterior distribution required for the auxiliary function, we introduce a training-free classifier capable of estimating pose categories in a plug-and-play fashion. Additionally, we employ various approximation techniques for terms of noisy states, significantly enhancing system performance. Our proposed solution, called RecDreamer, demonstrates a marked ability to mitigate the Multi-Janus problem, as confirmed by our experiments.