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

Forum topic · 小凯 · 2026-06-28

Summary

This paper introduces an RLHF-style approach for fine-tuning 3D-aware generative models directly on radiance-field density (sigma) values, without external meshes or shape priors. Unlike prior pipelines that optimize explicit surface representations via mesh conversion, the authors train a reward model on a small set of human preference samples—no reward pretraining required—and use it to fine-tune an unconditional 3D-aware face GAN (EG3D). The reward reads the continuous 3D density field of the neural radiance field (NeRF) directly, supplying a geometry-only learning signal that needs neither text conditioning, mesh extraction, nor multi-view rendering. A density-consistency constraint preserves 2D appearance while reshaping geometry, at a measurable but bounded distributional cost (FID-50k rising from 4.09 to 6.66). As a proof of concept, the generator fine-tuned from a single annotator's preferences produces face geometry preferred by users in 74.4% of pairwise comparisons. Paper by Archer Moore, Mingming Gong, and Liam Hodgkinson, arXiv:2606.27305 (June 2026).

Paper Overview

Field: Computer Vision Authors: Archer Moore, Mingming Gong, Liam Hodgkinson Published: 2026-06-25 arXiv: 2606.27305

Abstract

Reinforcement learning from human feedback (RLHF) for 3D generation is now established across a number of works, but most existing pipelines optimise explicit surface representations, often by converting radiance fields into meshes and training heavily on surface-supervised data. This work instead fine-tunes a pretrained 3D-aware generative model directly from a learned reward over radiance-field density (sigma) values, with no externally supplied mesh or shape prior.

Key Contributions

Mesh-free reward: The reward model reads the continuous 3D density field of a neural radiance field (NeRF) directly, requiring no pretraining and no external shape supervision.
Data efficiency: The reward trains easily on a small set of preference samples and yields robust improvements in 3D geometry.
Geometry-only signal: Applied to an unconditional 3D-aware face GAN (EG3D), the reward supplies a purely geometric learning signal—no text conditioning, mesh extraction, or multi-view rendering is needed.
Appearance preservation: A density-consistency constraint keeps the 2D appearance qualitatively similar while reshaping geometry, at a measurable but bounded distributional cost (FID-50k rises from 4.09 to 6.66).
Human evaluation: As a proof of concept, a generator fine-tuned from a single annotator's preferences produces face geometry preferred by users in 74.4% of pairwise comparisons.

--- *Automatically collected on 2026-06-28*