Suboptimal interpolation when training the detail predictor

edward1997104 / Wavelet-Generation

Neural Wavelet-domain Diffusion for 3D Shape Generation [SIGGRAPH Asia 2022]

MIT License

86 stars 12 forks source link

Suboptimal interpolation when training the detail predictor #7

Open MarkTension opened 1 year ago

MarkTension commented 1 year ago

Hello, thank you for sharing your code! I'd like to bring your attention to how F.interpolate is being used for training the detail predictor, in line: https://github.com/edward1997104/Wavelet-Generation/blob/main/trainer/trainer.py#L260C33-L260C46

The interpolation mode is not set, and defaults to 'linear'. Setting mode='trilinear' gives better results (because it uses the 3 spatial dimensions instead of 1) and gets rid of some of the artifacts that I see in the results. Cheers!

MarkTension commented 1 year ago

EDIT: Sorry, it defaults to 'nearest', but still I think 'trilinear' is more desirable.

edward1997104 commented 1 year ago

Oh... It is really nice for you to find out that as we used to consider this a design decision! We might test it out later to provide an updated model!