SAM3 segmentation node eats up vram

[Cloudreadypc]

I was using Wan2.2 low model to refine 4M/8M images. Three SAM3 segmentation nodes were used for mask. When device's set to GPU, the workflow finished in 70mins with RTX5070 (12GB) and 64GB ram. And quite often, I got OOM even at the third SAM3 node (First and second were fine).

When I set SAM3 device to CPU, no more OOM and the workflow finished in just 5mins.

I wonder if there's something wrong in the node, which consumes 3 times the vram needed. Thanks a lot.

Comfyui v0.4.0

[1038lab]

Yes, this behavior is possible.

SAM3 is very VRAM-heavy on GPU, especially when running multiple nodes on large images. When VRAM is close to full, CUDA memory pressure can cause OOMs and severe slowdowns, which is why CPU can be faster in this case.

This is expected behavior, not a bug.
As a workaround, try enabling unload_model, using fewer SAM3 nodes, or keeping SAM3 on CPU for very large images.