Manifoldsplat enables fast language guided edits of 3d head avatars
ManifoldSplat: Language-Guided Semantic Shape Editing of 3D Gaussian Head Avatars
Computer Vision and Pattern Recognition
Summary
Creating realistic 3D head avatars is possible, but changing their shape using simple language commands is usually slow and can mess up the avatar’s identity or movement. The authors present ManifoldSplat, a system that lets people quickly and precisely edit 3D avatar shapes by typing descriptions. It keeps the avatar’s look and animations intact by working within a structured shape space, and it can produce edited results in about 90 seconds on a regular computer. This method improves how closely edits match the language commands and keeps the avatar consistent when viewed from different angles.
What this means in practice
- •For game developers: Create customizable 3D head avatars that users can edit quickly using natural language without losing character identity or animation features.$Commercial implications: Enables interactive avatar customization tools for games, allowing players and designers fast, realistic, and user-friendly modifications.
- •For virtual reality content creators: Produce high-quality animatable 3D avatars with fine-grained shape edits based on text prompts, improving immersion in VR experiences.
Authors
Antonio Canela, Jordi Sànchez-Riera
Abstract
High-fidelity 3D head avatars have reached near-photorealistic quality. While recent methods enable text-driven manipulation, they struggle to provide fine-grained localized control, often entangling features or lacking geometric consistency. Modifying geometry through natural language currently requires slow per-prompt optimization or compromises identity and rigging. We present ManifoldSplat, the first end-toend framework for language-guided semantic shape editing of animatable 3D Gaussian Splatting avatars reconstructed from monocular videos. By performing edits within the structured FLAME manifold rather than directly optimizing an unstructured Gaussian cloud, we strictly preserve identity and animation. We introduce DeltaRegion, a per-region disentangled Conditional Variational Autoencoder (CVAE) delivering feedforward shape deltas, alongside a refining stage to recover view-consistent details. ManifoldSplat reconstructs and edits an avatar in ~90 seconds on a consumer GPU, rendering at ~800 FPS. Extensive evaluations demonstrate our approach sets a new state-of-the-art in localized prompt alignment, geometric coherence, and identity preservation. Project page and code: https://a-canela.github.io/manifoldsplat/