>>105631665
I'm still experimenting with this shit but:
1. Four steps is just way too little, even with self-forcing lora at full strength. I'm currently jumping back and forth between 6-8 steps.
2. Splitting your gen between two ksamplers, with the first one doing a couple of steps at high CFG helps out a lot.
3. lcm sucks, unipc is still king. ddim with ddim uniform is also pretty interesting.
4. Selfforcing lora by itself kills the motion, I'm trying to experiment by mixing it up with accvid and causvid at various weights. Hell, I also throw in the HPS/MPS rewarding loras at low weights for the first ksampler, which seems to help out a little.
5. Not only do you have to "boomer prompt" - you have to explain shit like you're talking to a toddler with a down syndrome. Every action, every movement has to be described as simply and at the same time as thoroughly as possible. Even then, I just can't make this one work