Reproduce “MorphAny3D: Unleashing the Power of Structured Latent in 3D Morphing” 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
Xiaokun Sun, Zeyu Cai, Hao Tang, Ying Tai, Jian Yang, Zhenyu Zhang
MorphAny3D introduces a training-free framework for high-quality 3D object morphing, enabling semantically consistent and temporally smooth transitions between diverse categories. It leverages Structured Latent (SLAT) representations with novel attention mechanisms and an orientation correction strategy, achieving state-of-the-art quantitative metrics with a FID of 111.95 and user preference of 86.73% for visual plausibility and smoothness.
Key insights
- Naive interpolation of key and value features in attention (KV-Fused) leads to semantic inconsistencies and localized artifacts, especially when blending 2D conditional features.
- Separately processing source and target information within cross-attention (MCA) is crucial for maintaining semantic plausibility during feature fusion.
- Explicitly leveraging temporal information from previously generated frames (TFSA) and correcting pose shifts (OC) are essential for achieving continuous and smooth 3D morphing sequences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.