Skimmed_CFG
A powerful anti-burn allowing much higher CFG scales for latent diffusion models (for ComfyUI)
CFG below at: 6/8/12/16/24/32, skimming scale at 6
Simply plug after the model loader (same for all the fours nodes):
Tested with SD3.5 turbo. To make it work set the skimming scale at 2~2.5
nodes:
Skimmed CFG: My version first version of this, works like a charm!
Skimmed CFG – replace: replace the values within the negative by those in the positive prediction, nullifying (actually giving an equivalent scale of 1) the effect of values targeted by the filter.
Skimmed CFG – linear interpolation: instead of replacing, does a linear interpolation in between the values. Highly recommanded!
Skimmed CFG – linear interpolation dual scales: Two scales. One named “positive” and one.. well “negative”. The name is more related to a visualy intuitive relation rather than fully from the predictions. A higher positive will tend to go towards high saturations and vice versa with the other slider.
Skimmed CFG – Difference CFG: Other algorithms based on changes depending on the scale. Brings back what goes too far in comparison.
Skimmed CFG – Timed flip: To be used with normal scales. Enhances the randomness and overall quality of the image. A bit less of an antiburn and a lot more of an enhancer. SDE Samplers react extremely well to it.
special option (first node only):
full_skim_negative: fully skim some part of the conflicting influence.
disable_flipping_filter: the skimming CFG will have much more control. It is meant to be used with the full_skim_negative toggle on. (the last image of this readme is an example)
Side-effects:
better prompt adherence
sharper images
less mess / more randomness due to less conflicts in between the positive and negative predictions
something something sometimes fused fingers with too low skimming CFG scale and too low amount of steps.
Tips:
The skimming scale is basically how much do you like them burned. 3 was the intended scale but suit to your needs.
The SDE samplers can still burn a little but much less
The SDE samplers can still totaly do nonsense with low steps (not with the node named “Timed flip”)
A too low skimming scale may require to do more steps
Recommanded skim: 2-3 for maximum antiburn, 5-7 for colorful/strong style. 4 is cruise scale.
a good negative prompt is a style negative prompt
to use super high scales it is not a bad idea to cut the negative before the end. You can find in this repository a node named “Support empty uncond”. Plug it after the skimmed cfg node. Then menu>advanced>conditioning>ConditioningSetTimestepRange and set the and at ~65%. This will avoid potential artifacts.
Pro tip:
It would be actually nice to have some support! because like this I will continue to share my findings!
Did you know that my first activity is to write creative model merging functions?
While the code is too much of a mess to be shared, I do expose and share my models. You can find them in this gallery! 😁
Other examples with a CFG at 100
Base image 🤭
Linear interpolation dual scale 10/0
Linear interpolation dual scale 5/0
Linear interpolation scale 5
Replace
Skimmed CFG node skimming scale at 4
Skimmed CFG node skimming scale at 4 with razor skim
full_skim_negative / disable_flipping_filter / skimming scale at 6 / CFG scale at 32
_do not lick the screen_
