HumanNOVA: Photorealistic, Universal, and Rapid 3D Human Avatar Modeling from a Single RGB Image

Forum topic · 小凯 · 2026-06-03

Summary

HumanNOVA is a new computer vision model that generates photorealistic 3D human avatars from a single RGB image in under one second, without test-time optimization. To overcome the scarcity of diverse, high-quality 3D human training data, the authors built a scalable data pipeline using two strategies: animating existing rigged assets with a wide range of everyday poses, and fitting existing multi-camera human captures to synthesize more diverse training views. These strategies scale the dataset to 100,000 assets, improving both quantity and diversity for robust training. Architecturally, HumanNOVA uses a feed-forward, token-conditioned avatar modeling framework: the input image and an estimated SMPL body (without detailed geometry or appearance) are encoded into compact token representations, which are fused via cross-attention to build a triplane-based 3D avatar representation. Experiments across multiple benchmarks show superior quantitative and qualitative results and robustness to diverse input image conditions. arXiv: 2506.00006; project page: https://HumanNOVA.github.io

Paper Overview

Field: Computer Vision (CV) Authors: Hezhen Hu, Wangbo Zhao, Lanqing Guo arXiv: 2506.00006

Introduction

This paper presents HumanNOVA, a photorealistic, universal, and rapid model for generating 3D human avatars from a single RGB image. Achieving both photorealism and generalization is challenging due to the scarcity of diverse, high-quality 3D human data.

Scalable Data Generation Pipeline

To address the data bottleneck, the authors build a scalable data pipeline with two strategies:

1. Animating rigged assets — leverage existing rigged 3D assets and animate them with extensive poses drawn from daily life. 2. Fitting multi-camera captures — utilize existing multi-camera human captures and employ fitting to generate more diverse views for training.

Together, these strategies scale the dataset up to 100k assets, significantly enhancing both the quantity and diversity of data for robust model training.

Architecture

HumanNOVA adopts a feed-forward, token-conditioned avatar modeling framework that runs in less than one second and requires no test-time optimization:

Given an input image and an estimated SMPL body (no detailed geometry or appearance needed), the model encodes both inputs into compact token representations.
These tokens serve as conditioning signals, fused via cross-attention to construct a triplane-based 3D avatar representation.

Results

Extensive experiments on multiple benchmarks demonstrate that the method outperforms existing approaches both quantitatively and qualitatively, and remains robust across diverse input image conditions.