TomePatchModel has virtually no effect on generation speed

Token Merging is reported to have a meaningful impact on performance as well as the memory footprint for the model.
Still, I have not been able to replicate these results.

Even generating images with high resolution, like 1024*1024 in a batch of 4 items on a SD1.5 checkpoint, the overall time being taken is basically the same, unless I set absurd thresholds like 90%, for which I only gain a 25% boost in speed.

Do you have any experience using it, or suggestions on what might be wrong?