search

google-research / multinerf

A Code Release for Mip-NeRF 360, Ref-NeRF, and RawNeRF
Apache License 2.0
3.57k stars 338 forks source link

Why the reconstructed image is very blurry #78

Open zerolfq opened 1 year ago

zerolfq commented 1 year ago

Why the reconstructed image is very blurry color_000

zerolfq commented 1 year ago

I execute according to the readme command, and the data set uses the data downloaded in mip Nerf360

jonbarron commented 1 year ago

How long did you train for?

zerolfq commented 1 year ago

How long did you train for?

250000 iterations，Use the configuration of 360.gin，Config.batch_size=16

jonbarron commented 1 year ago

The batch size needs to be way bigger.

Riga27527 commented 1 year ago

Hi, @jonbarron . I'm now using blender_refnerf.gin, but due to memory limitations, the batch_size is set to 2048, how many rounds of training would be better? Is 500,000 rounds enough?

jonbarron commented 1 year ago

I'd follow the "linear scaling rule" when changing batch size. You should be scaling up the number of iterations and scaling down the learning rate so that the model sees the same number of pixels as it had previously.