Reproduce “SketchFaceGS: Real-Time Sketch-Driven Face Editing and Generation with Gaussian Splatting” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bo Li, Jiahao Kang, Yubo Ma, Feng-Lin Liu, Bin Liu, Fang-Lue Zhang, Lin Gao

SketchFaceGS introduces the first unified, real-time framework for generating and interactively editing photorealistic 3D Gaussian head models directly from 2D sketches. It achieves editing latency of approximately 0.3 seconds and 243 FPS rendering, with an FID score of 44.60 for editing tasks and robust identity preservation (PSNR 31.12).

Key insights

  • A coarse-to-fine generative pipeline effectively translates sparse 2D sketches into dense, photorealistic 3D Gaussian head models by progressively resolving ambiguity.
  • Decoupling geometric and appearance information using parallel Transformer branches and an AdaIN-based alignment network is crucial for robustly handling diverse sketch and reference inputs.
  • Performing local edits through layer-by-layer feature fusion within the generative network's latent space ensures seamless blending, preserves unedited regions, and maintains 3D consistency more effectively than direct 3D compositing.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.