Existing score distillation methods are sensitive to classifier-free guidance (CFG) scale, manifested as over-smoothness or instability at small CFG scales, while over-saturation at large ones. To explain and analyze these issues, we revisit the derivation of Score Distillation Sampling (SDS) and decipher existing score distillation with the Wasserstein Generative Adversarial Network (WGAN) paradigm. With the WGAN paradigm, we find that existing score distillation either employs a fixed sub-optimal discriminator or conducts incomplete discriminator optimization, resulting in the scale-sensitive issue. We propose the Adversarial Score Distillation (ASD), which maintains an optimizable discriminator and updates it using the complete optimization objective. Experiments show that the proposed ASD performs favorably in 2D distillation and text-to-3D tasks against existing methods. Furthermore, to explore the generalization ability of our paradigm, we extend ASD to the image editing task, which achieves competitive results. The project page and code are at this link.

How to score the quality of the network output is an essential but long-neglected problem in DensePose, which dramatically limits the potential of the existing methods. To fill the blank in the quality estimation of DensePose, we conduct rigorous experiments to clarify the key factors that accurately reflect the quality of DensePose results. We find that the accurate results already exist in the candidate pool but are mistakenly removed due to the inappropriate quality scores. To solve this problem, we proposed DensePose Scoring RCNN (DS RCNN), a simple and comprehensive quality estimation framework to learn the calibrated quality score and select high-quality results from the pool. DS RCNN introduces a quality scoring module (QSM) and a quality perception module (QPM) into the existing high-performance pipeline. The QSM scores the quality of DensePose results by fusing diverse quality information, and the QPM enhances the ability of quality perception by extracting instance-aware quality features guided by the predicted IUV maps. Benefiting from the superiority of QSM and QPM, DS RCNN outperforms baselines by up to 4.8 AP on the DensePose-COCO dataset.

Data imbalance, a foundational problem in machine learning, has received little attention in DensePose and has become one of the main obstacles in front of existing methods. We reveal two imbalances in DensePose: inter-surface imbalance and intra-surface imbalance. First, the human body parts can be of various sizes, making 3D surfaces contain different numbers of annotations. Classifiers trained on such unbalanced data suffer a decline on surfaces with fewer annotations. Second, annotations within each surface are not uniformly distributed. Regressors trained on such uneven data suffer a decline in 3D surface areas with sparse annotations. To solve these imbalances, we propose PoiseNet which integrates adaptive equalization loss (AEQL) and block balanced localization (BBL). Specifically, to address the inter-surface imbalance, AEQL adaptively reweights surfaces based on the number of annotations and the classification scores. BBL alleviates the intra-surface imbalance by unevenly blocking each surface according to annotation distribution. Experimental results on the DensePose-COCO dataset show that our PoiseNet surpasses baselines by up to 1.4 AP.