Facing CUDA OOM

[Chaitaaa]

I am encountering consistent CUDA Out of Memory (OOM) errors when running ROSE locally on a GPU with 16GB VRAM (RTX 4070 Ti 16GB). Despite 16GB being sufficient for many video generation tasks, the default inference settings (resolution and frame count) appear too aggressive for this memory capacity. Could you please provide guidance on optimized settings for 16GB cards ?

[m1a0cX]

Sorry for late reply. Since our project mainly focuses on object and its side effect removal, we don't take the memory efficiency into our main consideration. To reduce memory usage, maybe you can try to use flash attention or other similar methods to reduce the memory usage and we welcome you to pull requests if you can. Also I will use my spare time to make the project more memory efficienct :)