ARG-Avatar

Animation-Rendering Guidance for Efficient Feedforward Avatar Reconstruction

Jianjin Xu, Soham Dasgupta, Fernando De la Torre

Carnegie Mellon University

Paper Code coming soon Demo coming soon

Given casual photographs or frames from a monocular video, ARG-Avatar reconstructs a canonical 3D Gaussian avatar in a single forward pass and animates it with any driving pose — with only 68M trainable parameters.

TL;DR

Feedforward avatar reconstruction doesn't need massive scale. ARG-Avatar is a lightweight model trained only on public data that turns a few photos into an animatable 3D avatar in ~2 seconds, outperforming methods that need far more training compute.

Performance of ARG-Avatar compared to baselines vs. trainable parameters and training compute. PSNR is calculated on the out-of-distribution NeuMan benchmark.

Interactive 3D Avatars

Canonical Gaussian avatars reconstructed from six input images of subjects not seen in training. Drag to rotate; pick a subject below.

Input images

Pose-Driven Animation

The same avatars driven by novel motion sequences.

Main Results

Novel pose synthesis with V input images. NeuMan is never seen in training; the other datasets use held-out validation subjects.

Method	V	NeuMan (novel poses)			Multiview Videos^*			3D Scans^**			HumanVid
Method	V	PSNR↑	SSIM↑	LPIPS↓	PSNR↑	SSIM↑	LPIPS↓	PSNR↑	SSIM↑	LPIPS↓	PSNR↑	SSIM↑	LPIPS↓
SIGMAN	1	19.62	0.957	0.066	16.52	0.944	0.076	15.27	0.908	0.112	16.17	0.862	0.162
IDOL	1	24.96	0.964	0.036	23.18	0.962	0.051	21.32	0.935	0.086	20.53	0.899	0.129
LHM	1	26.10	0.968	0.028	25.30	0.965	0.041	22.81	0.940	0.078	21.23	0.906	0.111
LHM++	1	27.71	0.970	0.023	25.34	0.966	0.039	22.31	0.939	0.078	22.03	0.910	0.104
	2	27.97	0.971	0.022	25.31	0.966	0.038	22.42	0.939	0.075	22.35	0.911	0.101
	4	28.03	0.972	0.022	25.60	0.966	0.037	22.57	0.940	0.074	22.47	0.911	0.099
	6	28.09	0.972	0.022	25.68	0.967	0.036	22.49	0.940	0.074	22.50	0.912	0.098
ARG-Avatar	1	27.02	0.966	0.024	25.86	0.962	0.034	24.09	0.937	0.058	21.64	0.897	0.100
	2	29.17	0.971	0.019	27.53	0.965	0.027	25.64	0.941	0.046	24.57	0.912	0.076
	4	30.14	0.973	0.017	28.53	0.967	0.024	26.59	0.944	0.040	25.44	0.916	0.069
	6	30.82	0.975	0.016	28.83	0.968	0.023	26.81	0.945	0.039	25.86	0.918	0.067

^*Multiview Videos combines DNARendering and 4D-Dress. ^**3D Scans averages 2K2K, THuman2.1 and Human4DiT, evaluated on novel views.
Baselines: SIGMAN, IDOL, LHM and LHM++; poses for all methods are estimated with LHM-Track.