Use this node like a RescaleCFG node, ... modelIn -> ThisNode -> ModelOut ... -> KSampler
See paper https://arxiv.org/pdf/2410.02416 for instructions about the other parameters. (Pages 20-21)
I have commented out the casting to .double() inside the function "project" as it was causing a parameter error on my machine (DirectML can't use 'double', I believe). Shouldn't cause any noticeable difference, probably.
Refresh your workflow and recreate the node if you're updating!
Edit5 - 29/11/2024:
Added "Guidance Limiter" : https://arxiv.org/abs/2404.07724 It drops the cfg (also, APG's functions) outside the specified range (ideally at the early steps and the later ones) improving variability and combating oversaturation. Starting and ending parameters set as per the paper suggestion, but may need further tweaking depending on model and other conditions.
If using CFGLimiterGuider from https://github.com/pamparamm/ComfyUI-ppm with same starting and ending parameters, there should be a speed boost, as the uncond is dropped. (It connects to SamplerCustomAdvanced)
Edit4 - 01/11/2024: Added an "adaptive_momentum". It gradually brings momentum towards 0 every step.
Adaptive_Momentum = 0 , No change to momentum
Adaptive_Momentum = 0.180 , Momentum will reach 0 roughly around the last step. I find that this helps to dramatically lower glitches/noise, specially on lower steps or higher CFG values. This is now the default value.
Adaptive_Momentum = 0.190 , Momentum will reach 0 roughly around half the steps.
Added a "print_data" boolean to print on console the momentum at every step.
Issues:
Using denoise < 1.0 will start momentum at a lower value if adaptive_momentum > 0.
Using denoise < 1.0 will NOT reset momentum_buffer between gens (changing resolution or a value on the node will). May cause reproducibility issues.
The denoise < 1.0 problem exists since the 1st version. I haven't found a way to adjust for these yet.
Edit3 - 21/10/2024: Now the Sampler's cfg is this node's previous "scale" value. May need to refresh the workflow and recreate the node.
Edit2: Fixed a bug where momentum_buffer.running_average wouldn't reset between gens, changed defaults based on my tests again.
If you tried changing resolution without triggering a re-patch, you'd see a nice error message...
Edit: Updated defaults, works well with high Norm_Threshold and higher scales