▪ フォーマットを統一したことで学習効率、安定性、忠実度が改善 関連研究:Direct3D-S2 Shuang Wu, et al., "Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention," in Proc. of NeurIPS 2025.
▪ 画像と3Dモデルの対応関係はCross Attentionでの学習に依存 →グローバルな意味情報を頼りに生成しがち Dong-Yang Li, et al., "Pixal3D: Pixel-Aligned 3D Generation from Images," in Proc. of SIGGRAPH 2026.
全特徴量を平均化 Back Projection Proj Injection: 特徴ボリュームを ノイズボリュームに加算 Dong-Yang Li, et al., "Pixal3D: Pixel-Aligned 3D Generation from Images," in Proc. of SIGGRAPH 2026.