Gorilla-Lab-SCUT / Fantasia3D

(ICCV2023) official repository for "Fantasia3D: Disentangling Geometry and Appearance for High-quality Text-to-3D Content Creation"
https://fantasia3d.github.io/
Apache License 2.0
711 stars 35 forks source link

memory cost #42

Open VLadImirluren opened 10 months ago

VLadImirluren commented 10 months ago

I haven't as large compute resource as yours. This make me hard to reproduce your work. Can you share some trick of reduce the memory cost but reach the same effect? (In brief, how to use lower batchsize but same result)

ARuiChen commented 10 months ago

1、Installing xformers can greatly reduce memory usage. 2、Use gradient accumulation technology that is easy to implement as illustrated in threestudio 3、https://github.com/threestudio-project/threestudio/blob/main/README.md?plain=1#L575 4、rent 5、I think 4 GPUs are enough to reproduce my work, but there should be some changes needed in the configuration. For geometry modeling, use the DMTet 128 resolution to achieve a balance between quality and memory consumption when limited to batch size. For appearance modeling, the addition of negative prompt can greatly improve the generation quality.

VLadImirluren commented 10 months ago

1、Installing xformers can greatly reduce memory usage. 2、Use gradient accumulation technology that is easy to implement as illustrated in threestudio 3、https://github.com/threestudio-project/threestudio/blob/main/README.md?plain=1#L575 4、rent 5、I think 4 GPUs are enough to reproduce my work, but there should be some changes needed in the configuration. For geometry modeling, use the DMTet 128 resolution to achieve a balance between quality and memory consumption when limited to batch size. For appearance modeling, the addition of negative prompt can greatly improve the generation quality.

thanks