Back to Main Page

We compare our method with different methods by removing each design component and showcase 8-view training results on cube.

Comparison with different methods

Burgert et al. [1]

Baseline

Random Patch

w/o resolution scaling

w/o resolution scaling

w/o camera jitter

Ours

Burgert et al. [1]

Baseline

Random Patch

w/o resolution scaling

w/o resolution scaling

w/o camera jitter

Ours

Comparison Table

Camera jitter × × × random scheduled × scheduled
Patch denoising × × ✓ ✓ ✓ ✓ ✓
Resolution scaling × × × × × ✓ ✓

More Views

8 views of 8 corners, each face is shared with 4 prompts. (random selected prompts)

Failure Cases

Duplicate Objects

Blending objects

Less natural objects

Easy to recognize face,
human nature ↑ back to top