You likely know more than me about doing it, but this is my source

https://wiki.archlinux.org/title/QEMU/Guest_graphics_acceleration

Single GPU passthrough

Currently, PCI passthrough works for dual-graphic cards only. However, there is a workaround for passing a single graphic card. The problem with this approach is that you have to deattach the graphics card from the host and use ssh to control the host from the guest.

When you start the virtual machine, all your GUI apps will be force terminated. However, as a workaround, you can use Xpra to detach to another Display before starting the virtual machine and reattach the Apps to display after shutting down the virtual machine.

If you have NVIDIA GPU, you may need to dump your GPU's vBIOS using nvflashAUR and patch it using vBIOS Patcher.

NVIDIA vGPU

By default, NVIDIA disabled the vGPU for consumer series (if you own an enterprise card go ahead). However, you can unlock vGPU for your consumer card.

You will also need a vGPU license, though there are some workarounds.

Follow this guide to manually setup a Windows 10 guest with NVIDIA vGPU.

Once I got my virtualization settings set up correctly in UEFI, and KVM was my hypervisor instead of QEMU TCG, my performance did seem pretty good. Maybe it's just working correctly without having to follow these steps?