liuyuan-pal / SyncDreamer

[ICLR 2024 Spotlight] SyncDreamer: Generating Multiview-consistent Images from a Single-view Image
https://liuyuan-pal.github.io/SyncDreamer/
MIT License
906 stars 39 forks source link

training vram limitation #56

Open immaturehearts opened 11 months ago

immaturehearts commented 11 months ago

how can i train this model on a single 16G GPU, please?

liuyuan-pal commented 11 months ago

It is not feasible to train the SyncDreamer on a 16G GPU because it needs a large batch size and the parameters already consumed a large amount of VRAM.