Closed bitinn closed 5 years ago
The released version is purely implemented on the CPU with a single core rather than on the GPU.
The released version is purely implemented on the CPU with a single core rather than on the GPU.
QuadriFlow internally calls the procedures in Instant Mesh. So it is expected that it will be slower than IM without GPU acceleration.
Thx both for the explanation.
Perhaps I am missing something, but my manual benchmark suggests Instant Meshes somehow run much faster (5x-10x) than QuadriFlow:
test3.obj
is3bommes/Buddha_input.obj
, I have ran a few similar tests, on macOS they are consistently slower than Instant Meshes.In the center is Instant Meshes' result, Left/Right are QuadriFlow's (Right:
-adaptive
enabled; Left:-adaptive
disabled), all 3 results have tris count at around 15K.(
-mcf
and-sat
flags increases the compute time further, albeit with better results, so I omit them here.)Is there an obvious reason as to why this contradicts with the paper's test results?
(The key difference here might be that I am on Intel GPU; but CPU speed is comparable.)