Rendering Clouds in Redshift (11GB memory limited GPU)

[Activate]

I am rendering a small animation which includes a camera moving through clouds. Here is a first previz:

I am wondering if I could get advice on my approach. I have a few hundred clouds overall but only five clouds that are directly seen by the camera over 240 frames of animation.

- I wrote a small solver to identify which clouds enter the view of the camera, e.g. by mapping points to NDC coordinates.

- Non visible clouds are created with the cloud node and a small uniform sampling rate of 50.

- Visible clouds use 1200 uniform sampling. (Video resolution will be 1920x1080). Each cloud takes between 1-2 GB of memory.

At first I created an RS Proxy for each cloud but that used too much memory. To deal with this, I took a box and mapped it's normal coordinates to NDC to give me the camera frustrum which I turned into a vdb. For each frame of the animation, I intersected the frustrum vdb with the vdbs of my five clouds and saved that off as a Redshift proxy:

Each proxy is ~250MB per frame. The renders are also quite noisy but Optix is doing a reasonable job at denoising them:

Bottom half is denoised - top half with noise.

The clouds still look somewhat unnatural which may require some more tweaking.

My question here really is if this is the overall correct approach for dealing with the memory limitations of GPUs.

Thank you.

[bunker]

you can try reducing the voxel resolution based on camera depth:
also clouds are not that sharp in real life, and with that camera movement you should have more motionblur 


reduce_voxelres_by_camera_depth_bunker.hiplc

[Activate]

Thank you for that suggestion and the example. I will try that.

As for motion blur, I have not figured that out yet. I was going to try to add it in Nuke. As far as I can tell, Redshift does not produce motion vectors for volumes but can certainly produce motion blur directly.