SplatCtrlA: Generalizable Single Image to Fully
Controllable 3D Avatar

ECCV 2026

Jun Xiang1,
Yudong Guo1,
Boyang Guo1,
Yancheng Yuan2,
Juyong Zhang1
1University of Science and Technology of China, 2The Hong Kong Polytechnic University
Paper (Uploading) arXiv (Uploading) Code (coming soon)

Our method reconstructs a 3D human avatar from a single image within seconds through a single feed-forward pass, supporting natural control over body pose, gestures, and facial expressions.

Abstract

Expressive human avatar creation from a single image is highly challenging due to the inherently ill-posed nature of the problem, as well as the complexities of appearance preservation and dynamic human modeling. To address these challenges, this work presents a comprehensive pipeline for training a large feed-forward model that efficiently generates fully controllable 3D avatars from single-image. Due to the scarcity of consistent training data, we construct a large-scale 3D Gaussian avatar dataset to support model training. To better recover appearance details, we propose an input-aware decoding scheme that fully leverages information from the input image. Furthermore, to achieve comprehensive full-body control, we introduce a Gaussian blending–based facial enhancement module and apply Gaussian geometric constraints to stabilize expressive avatar generation.

Method

The input image is processed in two parallel streams: together with the cropped head image, it is fed into a transformer-based image encoder to extract UV feature maps; meanwhile, via the proposed input-aware scheme, it is explicitly projected into the SMPL-X UV space and combined with semantic information to guide the separate decoding of Gaussian geometry and color attributes. The Inputaware Decoder directly predicts the Gaussian UV attribute maps, serving as a 2D representation of the 3D human avatar in the canonical space, which can then be deformed to the target pose via Linear Blend Skinning (LBS) for free animation.

Novel View Synthesis

Free Pose Animation

Comparison

Generalizability

BibTeX

@inproceedings{xiang2026splatctrla,
  title={SplatCtrlA: Generalizable Single Image to Fully Controllable 3D Avatar},
  author={Jun Xiang and Yudong Guo and Boyang Guo and Yancheng Yuan and Juyong Zhang},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2026}
}