https://huggingface.co/ByteDance/Hyper-SD/
tl;dr figure it out yourself scrub.
No longer downsized! Inferior to optimized checkpoint merges.
These are just suggestions. Experimentation is required and you will waste some credits to understand this LoRA.
Weight: 0.8 - 1.0
Sampling method: Anything! Some models react very badly with certain samplers! I recommend either Euler, DDIM or Euler A.
The versions are NOT representative of number of steps you use! 2steps does NOT mean you should use only 2 steps!
Sampling steps for 1step, 2steps: 4 minimum.
Sampling steps for 4steps: 5 minimum.
Sampling steps for 8steps: 8 minimum.
Sampling steps for cod-4steps-CFG: 8 minimum.
Good sampling steps depends on the model that you use. A good number to start with is 7, then use higher steps or lower steps.
CFG is usually around 1.0-1.8. It is completely dependent on which model that you use. A good number to start with is 1.5, then use higher or lower CFG
cod-4steps-CFG works with higher CFG, meaning you get to use CFG meaningfully as opposed to extremely low values. It still is dependant on the model. A good number to start with is 4.0, then use higher or lower CFG. For realistic models that often has lower CFG values, start with 2.0 instead.
TODO: rentry articles going in-depth about hyper-sd capabilities and shortcomings.
Cover artwork: https://pixai.art/artwork/1785851893719513342
