Boost generation speed by using this lora instead of real CFG. Use BasicGuider or set CFG=1.0
In the examples, top: real cfg only, bottom: split schedule between real cfg and distilled (except the fox, which is middle: distilled only, bottom: no cfg)
It can be used for the full generation schedule, but composition and quality may be worse than real CFG, so for better results, split the schedule to use real CFG without the lora for the first 10-30% of steps, then CFG=1 with the lora for the remainder of steps. Example workflow included in "training data"
Works well with gradient_estimation sampler.