Sculpting NeRF Geometry: Human-Preference Fine-Tuning of a 3D-Aware Face GAN Without Meshes

Forum topic · 小凯 · 2026-06-28

Summary

A 2026 arXiv paper (2606.27305) by Archer Moore, Mingming Gong, and Liam Hodgkinson introduces an RLHF-style fine-tuning method for 3D-aware generative models that operates directly on neural radiance field (NeRF) density (sigma) values, rather than optimizing explicit surface representations such as meshes. The reward model requires no pretraining, trains easily on a small set of human preference samples, and needs no external mesh or shape prior. Applied to the unconditional 3D-aware face GAN EG3D, the learned reward reads the continuous 3D density field directly, providing a geometry-only signal that requires no text conditioning, mesh extraction, or 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, a generator fine-tuned from a single annotator's preferences produced face geometry preferred by users in 74.4% of pairwise comparisons.

Paper Overview

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

Abstract (translated from the Chinese summary)

Reinforcement learning from human feedback (RLHF) for 3D generation is now established across a number of works, but most existing pipelines optimize 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.

The reward model requires no pretraining, trains easily on a small set of preference samples, and yields robust improvement in 3D geometry. Working on an unconditional 3D-aware face GAN (EG3D), the reward reads the continuous 3D density field of the neural radiance field (NeRF) directly and supplies a geometry-only learning signal — requiring neither text conditioning, mesh extraction, nor multi-view rendering.

A density-consistency constraint preserves qualitatively similar 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, a generator fine-tuned from a single annotator's preferences produces face geometry preferred by users in 74.4% of pairwise comparisons.

Original Abstract (excerpt)

> 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. We instead fine-tune 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. The reward model requires no pretraining, trains easily on a small set of preference samples, and yields robust improvement in 3D geometry. Working on an unconditional 3D-aware face GAN (EG3D), our reward reads the continuous 3D density field of the neural radiance field (NeRF) directly and supplies a geometry-only learning signal, r...

---

*Auto-collected on 2026-06-28*