akanimax / akanimax/thr3ed_atom
Triplane - MLP needs regularization
未關閉
- 主要語言
- Python
- 星號
- 115
- 分支
- 9
- PR 合併指標
- 30 天內沒有已合併 PR
描述
The current triplane-mlp implementation, without any regularization, has decent performance, but shows density floater artifacts. Attached qualitative on Hotdog.
Hotdog results:
```
Mean PSNR on holdout set: 32.38438158842274
Mean LPIPS on holdout set: 0.09272133961319923
```
https://user-images.githubusercontent.com/16590868/196246830-3d055bc0-ac0f-440c-8fcb-2105cf83f958.mp4
Mic results:
```
Mean PSNR on holdout set: 29.51383826194816
Mean LPIPS on holdout set: 0.08450011044740677
```
https://user-images.githubusercontent.com/16590868/196247136-476e5ca2-83ec-449d-ad09-24d2bbaee905.mp4
Need to improve these results before merging in master.
貢獻指南
這個儲存庫沒有索引到貢獻指南
評估
這個 Issue 還沒有評估資料。