We ablate on the design choice of forward LBS and inverse LBS. We compare the animation results of models trained with forward LBS (left) and with inverse LBS (right) strategies. When using forward LBS training, certain floating artifacts can be observed around the hip and shoulder regions during the animation process.
We compare avatar results with different depth / normal conditions.
We update avatar results without prefixes such as "8K, HD, studio-like, blender, ultra-realistic, unreal" .
After fixing this bug, we reduce the "floater" artifacts, producing cleaner and artifact-free outputs. .
The original video results of text-driven animation. When the character moves away from the camera ("zoom-out"), the character's body only occupies a smaller pixel area, which can reduce visual clarity.
We compare AvatarStudio with current SoTA methods En3D, HumanGaussian, Rodin.
HumanGaussian
En3D
Rodin
Ours
a woman wearing ski clothes
a Texas ranger
We provide more examples of CFG-rescale and part super-resolution. The additional examples clearly illustrate how Part-aware SR contributes to detail refinement in specific body parts, enhancing overall visual fidelity. Moreover, we conduct further tests for CFG-rescale across several examples. These tests demonstrate that CFG-rescale effectively mitigates over-saturation issues, resulting in more natural and visually pleasing outputs.
@article{anon2023avatarstudio,
author = {Anonymous},
title = {AvatarStudio: High-fidelity and Animatable 3D Avatar Creation from Text},
joural = {OpenReview},
year = {2023},
}