Sculpting NeRF Geometry: Human-Preference Fine-Tuning of a 3D-Aware Face GAN

While 3D-aware generative models like EG3D can produce photorealistic 2D images, the underlying 3D geometry often suffers from artifacts such as irregular surfaces, bumps, and discontinuities. The authors investigate whether these geometric defects can be corrected by fine-tuning the model using human preferences, specifically avoiding the common reliance on explicit 3D mesh priors or text-based prompts.

The researchers developed a reward model that operates directly on the implicit 3D density field (sigma values) of a pretrained NeRF. By collecting a dataset of 4,346 pairwise comparisons from a single annotator, they trained a reward model to score the quality of these density fields. This reward signal is then used to fine-tune the generator's parameters. To ensure the 2D appearance of the generated faces remains consistent, they applied a density-consistency constraint during the fine-tuning process.

The proposed method successfully improves the geometric quality of the generated faces, particularly in areas like the nose and sides of the face, which are common failure points in existing models. In a blind user study, the fine-tuned geometries were preferred by participants in 74.4% of pairwise comparisons compared to the original model outputs. The authors demonstrate that this approach is computationally efficient, requiring only a few hours of training on a single GPU, and avoids the "sink trap" issues often encountered in text-conditioned 3D generation.

This work provides a streamlined, prompt-free alternative to existing 3D reinforcement learning from human feedback (RLHF) pipelines. By working directly with the implicit density representation, the method bypasses the need for complex mesh extraction or multi-view rendering during the training loop. This suggests that high-quality 3D geometry can be achieved by focusing on the underlying volumetric representation, offering a more direct path to improving 3D generative models.