I’m only able to benchmark it against CPU lightmass because I’ve never used the unofficial GPULM plugin. I also never run it in slow mode because that takes forever. Compared to a CPU bake, it’s nothing short of amazing as far as speed is concerned.
As for quality, it could be better. I get some hue-shift artifacts that are noticeable on smooth surfaces and the VLM seems noisy though perhaps I am just not using enough samples. I’ve not seen any negative VLM values though.
My project is reliant on low indirect lighting intensity and low MinBrickError for VLMs (as I mainly use them for occlusion, not indirect light) though, so it’s important to me at least that it reaches some degree of feature parity with CPU lightmass.