Kailu Wu

Results 20 comments of Kailu Wu

理论上,3060应该只需要不到2分钟。根据输出信息来看,ONNX 并没有运行在GPU上,而是CPU上。所有超分辨率操作会调用 onnx 进行运行,这个在CPU会耗费极大量时间,GPU则不会。请检查 onnxruntime-gpu 是否正确安装

> > 理论上,3060应该只需要不到2分钟。根据输出信息来看,ONNX 并没有运行在GPU上,而是CPU上。所有超分辨率操作会调用 onnx 进行运行,这个在CPU会耗费极大量时间,GPU则不会。请检查 onnxruntime-gpu 是否正确安装 > > 昨天看到消息,折腾了一整天,还是不行,安装onnxruntime-gpu包后,import onnxruntime onnxruntime.get_available_providers() 显示 ['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']。onnxruntime.get_device() 显示gpu。但是一旦generate 3d就会提示 onnxruntime::ProviderLibrary::Get [ONNXRuntimeError] : 1 : FAIL : LoadLibrary failed with error...

onnx 这块的任务是完成一个x4 的超分辨率(从512->2048)占的显存会比较大,我们自己测试是至少要8GB显存(6GB都不行),加上别的,应该是18GB的样子。如果要压缩到12GB,要么这块全跑在CPU上(就十五分钟一次,巨慢),要么去掉这个x4的超分,改成普通的resize(理论上不会差太多)。 https://github.com/AiuniAI/Unique3D/blob/4263b8c836950babedd9c8b6769aa7c41afa9dbc/scripts/refine_lr_to_sr.py#L53 这里直接把 img 转成 PIL Image 然后调用 PIL Image 的 resize 函数,增加到原分辨率的四倍(长宽各4)。结果转成np.ndarray 作为 output。 至于运行时卡顿的话,这个要看看你电脑内存够不够了,大概30GB的内存还是需要的(主要是模型太多了)。

It looks like the nvdiffrast library is not installed correctly. You can try the following installation script (with root privileges): ``` # Setup nvdiffrast apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install...

I didn't understand what is meant by ‘custom input images’? If you want to use 3D generation as a function call, our suggestion is to use Gradio Client to interact...

We use multiview high resolution and low resolution pairs. Multiview images comes from blender's rendering results for the objaverse dataset.

Yes, we use a (256,512) resolution pair for the first stage of super-resolution training, where the 256 resolution portion is augmented using downsampling to a random resolution and then upsampled...

运行显存需求和输入图像分辨率基本上没有关系(即2048分辨率和240分辨率用的是差不多的) 通过在各种 pipeline 增加 enable_model_cpu_offload 可以降低到15GB 显存需求左右,但是会导致内存泄漏(我也不知道为什么)

Theoretically any diffusion inference acceleration work can be used, we don't have enough time to do acceleration for inference. If you can help, we'd really appreciate it!

We recommend upgrading the OpenGL version or just replacing ***all*** dr.RasterizeGLContext with dr.RasterizeCudaContext (it will be a bit slower, but no longer depends on OpenGL).