SeparateGen: Semantic Component-Based 3D Character Generation From Single Images.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Generating detailed 3D characters from single images is difficult due to semantic component entanglement.
• Existing methods often result in poor mesh topology, limiting usability in animation and rigging.

Purpose of the Study:

• Introduce SeparateGen, a framework for high-quality 3D character generation from single images.
• Explicitly reconstruct characters into distinct semantic components (body, clothing, hair, shoes).

Main Methods:

• Utilize a multi-view diffusion model for consistent canonical A-pose image generation.
• Employ a novel component-aware reconstruction model (SC-LRM) for high-fidelity component decomposition and reconstruction.
• Introduce SC-Anime, a large-scale dataset of 7,580 anime-style 3D characters with component-level annotations.

Main Results:

• SeparateGen significantly outperforms state-of-the-art methods in reconstruction quality and multi-view consistency.
• The component-based approach effectively resolves mesh entanglement issues.
• Generated characters are suitable for seamless rigging and asset reuse.

Conclusions:

• SeparateGen advances the generation of application-ready 3D characters from single images.
• The SC-Anime dataset and code will be publicly released to facilitate further research.