It depends on the model but I’ve seen image generators range from 8.6 wH per image to over 100 wH per image. Parameter count and quantization make a huge difference there. Regardless, even at 10 wH per image that’s not nothing, especially given that most ML image generation workflows involve batch generation of 9 or 10 images. It’s several orders of magnitude less energy intensive than training and fine tuning, but it is not nothing by any means.

The training is a huge power sink, but so is inference (I.e. generating the images). You are absolutely spinning up a bunch of silicon that’s sucking back hundreds of watts with each image that’s output, on top of the impacts of training the model.

That’s not the case for the newer open source drivers from nvidia. They’re only compatible with the last few generations of cards but they’re performant and the only feature they lack is CUDA to my knowledge. Not talking nouveau here