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 feed-forward model that generates photorealistic, generalizable 3D human avatars from a single RGB image in under one second, with no test-time optimization. The authors address the scarcity of diverse, high-quality 3D human training data through a scalable data generation pipeline combining two strategies: animating existing rigged assets with extensive everyday-life poses, and using fitting on multi-camera human captures to synthesize more diverse training views. This scales the dataset to 100k assets, boosting quantity and diversity. Architecturally, HumanNOVA encodes the input image and an estimated SMPL body mesh into compact token representations, which are fused via cross-attention to construct a tri-plane-based 3D avatar representation, without requiring detailed geometry or appearance supervision. Experiments across multiple benchmarks show superior quantitative and qualitative results and robustness to diverse input images. Paper: arXiv:2506.00006; project page: https://HumanNOVA.github.io

Overview

Field: Computer Vision (CV) Authors: Hezhen Hu, Wangbo Zhao, Lanqing Guo Published: 2026-06-03 arXiv: 2506.00006

Abstract

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. To address this, the authors build a scalable data generation pipeline that follows two strategies:

1. Leverage existing rigged assets and animate them with extensive poses from daily life. 2. Utilize existing multi-camera captures of humans and employ fitting to generate more diverse views for training.

These two strategies enable scaling up to 100k assets, significantly enhancing both the quantity and the diversity of data for robust model training.

Architecture

HumanNOVA adopts a feed-forward, token-conditioned avatar modeling framework that enables rapid inference in under one second, without test-time optimization. Given an input image and an estimated SMPL body mesh — without detailed geometry or appearance — the model:

Encodes both inputs into compact token representations.
Uses these tokens as conditioning signals, fused via cross-attention to construct a tri-plane-based 3D avatar representation.

Results

Extensive experiments on multiple benchmarks demonstrate the method's superiority both quantitatively and qualitatively, with robustness across diverse input image conditions.

Project page: https://HumanNOVA.github.io