Go to the Extensions tab, then Install from URL, use the URL for this repository.
Pick methods.
Delaying the start can be beneficial, as can early ending. This sort of correction has a tendency to prevent extremes of lighting. custom functions:
- M: mean
- m: median
- q(n): quantile. 0.5 is same as median; using high values will darken the image appropriately for sunsets, etc.
- rM(n, m): mean of range, rM(0, 0.5) gives mean of lowest 50%
- c: channel index, will be 0, 1, 2 or 3. Example uses: q([0.9, 0.5, 0.5, 0.7][c]); [m, 0.0, 0.0, 0.1][c]
- generation parameters now also saved to params.txt
larger update:
- added local average to overall, can use as a local contrast adjustment, adjustable blur radius
- added strength sliders for per channel and overall (removes previously hidden multipliers, defaults match those previous values)
- minor UI reshuffle
fixed bug with centered mean and topK of 0.5. (0.5-0.5)=0, who knew?
added saving/loading of custom functions
Public domain. Unlicense. Free to a good home. All terrible code is my own. Use at your own risk, read the code.
General idea from (Birch Labs)[https://birchlabs.co.uk/machine-learning#combating-mean-drift-in-cfg] but this is after CFG
SoftClamp method by (Timothy Alexis Vass)[https://huggingface.co/blog/TimothyAlexisVass/explaining-the-sdxl-latent-space]