Core Concepts
EpiDiff efficiently generates multiview-consistent images using epipolar constraints, improving quality and diversity.
Abstract
EpiDiff introduces a localized interactive multiview diffusion model.
Utilizes epipolar attention block for cross-view interaction.
Enhances consistency and quality in multiview images.
Outperforms previous methods in speed and quality metrics.
Improves reconstruction from generated multiviews.
Stats
EpiDiffはわずか12秒で16のマルチビュー画像を生成します。
EpiDiffはPSNR、SSIM、LPIPSなどの品質評価メトリクスで以前の方法を上回ります。