Summary:
- Higher the scale the more refined it becomes (Ex. better hands and better limbs)
- Lower the scale the faster the generation and can use bigger batch sizes
- When using high scale, scale the CFG higher to have more adherence to prompt, otherwise it will not adhere to it
- I normally use lower CFG + Sampling then use a highres fix to save time while generatingYou can also use the image enhancer colab at the end of this document to resize and upscale the images
Conclusion:
Stick to 20-28 when testing, when ready to execute for display use 50. 100-150 too high but still usable (I just don’t recommend using it)
Summary

20

28

50

100

150
Sampling Steps 20
Best out of the batches:
Pro:
- Fast generation (Even in batches of 8)
Cons:
- Bad hands (more than 5 fingers, or blends into background)
- Limbs are often replicated (extra hand, extra arms)
Sampling Steps 28
Pro: