We evaluate the quality of novel view synthesis for different methods, presenting quantitative results in Table 2, and qualitative results in Figure 3.
Zero123 produces visually reasonable images, but they lack multi-view consistency since it operates on each view independently.
Our method is capable of generating images that exhibit semantic consistency with the input image and maintain high consistency across multiple views.
To validate our proposed cross-domain diffusion scheme, we explore various settings, comparing cross-domain attention's effectiveness with other approaches.
Collection
[
|
...
]