search

szymanowiczs / splatter-image

Official implementation of `Splatter Image: Ultra-Fast Single-View 3D Reconstruction' CVPR 2024
https://szymanowiczs.github.io/splatter-image
BSD 3-Clause "New" or "Revised" License
795 stars 54 forks source link

About the training time #50

Closed aruiplex closed 2 months ago

aruiplex commented 3 months ago

Hi! I am reproducing your fantastic work. I am training the SRN car and SRN chair datasets. It needs 55h in stage 1 with RTX4090+7950x. Is it normal training time-consuming? Could you please provide more information about training stage 1 and stage 2? The description in the readme is a little confusing to me and the paper is not include this part.

Thx in advance.

szymanowiczs commented 2 months ago

55h sounds about right, that’s 2.5 days - the paper says total for 1+2 should take around 7.

Stage 1 does most of the work, stage 2 includes the LPIPS loss which makes the reconstructions sharper, although training takes a bit longer because you need to propagate through the VGG network. Hope this helps.