ComfyUI Node
PowerLawCFG
Implements 'Power-law CFG' as found in the same paper (2502.07849). Disables the speed optimization when `cfg=1`, as the negative term no longer cancels out.
PowerLawCFG
- model
- MODEL
◄alpha0.90►
◄parameterizationscore►
◄print_debugfalse►
Categoryadvanced/guidance
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| alpha | FLOAT | 0.90-0.99–100 | The exponent for the 'power law,' which finds a number to multiply `cfg` by at each step. Values > 0 are recommended by the paper authors. 0: disable <0: guidance which speeds up convergence to the target at early times >0: dampen guidance if positive and negative are similar; amplify guidance if positive and negative are dissimilar. |
| parameterization | COMBO | score | The space in which to calculate the scale. You do NOT have to match this with your model type, they can be different. In practice, `score` seems to be the most stable. score: The default implementation. x0: 'Rescaled Power-law CFG' in the paper. eps/v/flow: Other parameterizations for completeness's sake. |
| print_debug | BOOLEAN | false | Print this node's calculations to console at each sampling step. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |