use scikit instead of torchmcubes - #26
Conversation
Just out of interest are there any obvious quality benefits over using scikit-image or is this mostly just an installing and testing solution? Speed benefits, wider range of GPU support etc? It would be interesting to actually see a full comparison between the two. |
|
@mrbid speed wasn't impacted in my tests. The issue with torchmcubes is that it lacks prebuilt wheels and has a dependency issue with torch that makes it harder to install |
|
@flowtyone how did you actually solved the problem? If you share a step by step would help a lot. I am struggling a lot to make it work |
|
@math-sasso this PR works, just use my fork branch, it has these changes |
|
https://github.com/flowtyone/TripoSR/commit/341657e34f98f314cdfdcfbabe689d0d52dde7a5 It's quite a minimal and elegant patch. No need to even fork it tbh. This should be merged as an optional function in TripoSR, allowing the user to select which is used - ofc that wont solve the dependency issue, and I don't know if there is a similar ifdef system in Python as there is in C. |
|
I have an error with your fork: I was going to bench it against the original and this modified version by @thatname here: #22 (comment) |
|
@mrbid @flowtyone Works on ROCM with this change. |
Increase usability by using scikit-image instead of the old torchmcubes. This makes the project much easier to install and test.