Avatar Concept Slider: Controllable Editing of Concepts in 3D Human Avatars

TL;DR

The paper addresses the challenge of precisely editing high-level semantic concepts in 3D human avatars given the ambiguity of natural language prompts. It introduces Avatar Concept Slider (ACS), which combines a Linear Discriminant Analysis–driven concept axis with a Concept Sliding Loss, a PCA-based Attribute Preserving Loss, and a concept-sensitive 3DGS primitive selection to enable slider-like editing within the SDS optimization framework. Key contributions include learning a discriminative concept axis $oldsymbol{b}_c$, disentangling edits from identity attributes via PCA bases, and efficiently editing only the most concept-sensitive primitives, all integrated with a diffusion-prior fine-tuning via LoRA adapters. The result is controllable, high-fidelity avatar editing that preserves identity, with demonstrated improvements over text-driven baselines and notable gains in editing efficiency for large 3D Gaussian Splatting representations.

Abstract

Text-based editing of 3D human avatars to precisely match user requirements is challenging due to the inherent ambiguity and limited expressiveness of natural language. To overcome this, we propose the Avatar Concept Slider (ACS), a 3D avatar editing method that allows precise editing of semantic concepts in human avatars towards a specified intermediate point between two extremes of concepts, akin to moving a knob along a slider track. To achieve this, our ACS has three designs: Firstly, a Concept Sliding Loss based on linear discriminant analysis to pinpoint the concept-specific axes for precise editing. Secondly, an Attribute Preserving Loss based on principal component analysis for improved preservation of avatar identity during editing. We further propose a 3D Gaussian Splatting primitive selection mechanism based on concept-sensitivity, which updates only the primitives that are the most sensitive to our target concept, to improve efficiency. Results demonstrate that our ACS enables controllable 3D avatar editing, without compromising the avatar quality or its identifying attributes.

Figures (8)

• Figure 1: Illustration of our controllable concept editing. In this example, the user has provided text descriptions for two opposing ends of a concept: 'short sleeves' vs 'long sleeves'. Our ACS allows users to specify the exact level of concept expression (e.g., sleeve length) that is desired, by moving the knob on the slider bar.
• Figure 2: (a) Overview of the fine-tuning stage, where an adapter is fine-tuned to learn slider-like capabilities. Firstly, using the provided descriptions of the positive ($c_p$) and negative ($c_n$) side of the concept, we extract the corresponding positive and negative features. Then, to fine-tune the adapter $\Delta \phi$, our proposed Concept Sliding Loss (Sec. \ref{['Sec:ConceptAnalysis']}) is applied to learn the ability to slide across opposing ends of the target concept. The attribute-preserving loss (Sec. \ref{['sec:identity_analysis']}) helps to retain the key identifying attributes of the avatar. (b) After fine-tuning, the adapter is applied in an SDS optimization pipeline to achieve controllable concept-specific 3D avatar editing. The proposed 3DGS primitive selection mechanism (Sec. \ref{['Sec:Primitive_Selection']}) further improves efficiency, by optimizing only the most related primitives to the target concept.
• Figure 2: Efficiency comparisons of our method.
• Figure 3: Our editing results using various concept sliders.
• Figure 4: Evaluation of our method's precise controllability for several concepts. We present the avatars generated across smaller steps (i.e., in steps of 0.5) along the concept sliders.