sigmas_tools_and_the_golden_scheduler
A few nodes to mix sigmas and a custom scheduler that uses phi, then one using eval() to be able to schedule with custom formulas.
Nodes (12)
The 'Align Your Steps' schedule, without the guesswork
A scheduler shaped like the tail of a bell curve
Turn a schedule into one number your noise nodes can use
See your schedule before you burn the GPU on it
A calculator for your noise schedule — type the formula, get SIGMAS
Blend 25 schedules into one, all at once
The lazy way to split the difference between two schedulers
Start as one scheduler, end as another
One number that makes every step hit harder (or softer)
The exact noise range your model actually lives in
Take the opening from one scheduler, the ending from another
The golden ratio as a noise schedule — and why it kind of works
sigmas_tools_and_the_golden_scheduler
A few nodes to mix sigmas and a custom scheduler that uses phi, then one using eval() to be able to schedule with custom formulas.
Nodes
Merge sigmas by average: takes sigmas_1 and sigmas_2 as an input and merge them with a custom weight.
Merge sigmas gradually : takes sigmas_1 and sigmas_2 as an input and merge them by starting with sigmas_1 times the weight and sigmas_2 times 1-the weight, like if you want to start with karras and end with simple.
Multiply sigmas: simply multiply the sigmas by what you want.
Split and concatenate sigmas: takes sigmas_1 and sigmas_2 as an input and merge them by starting with sigmas_1 until the chosen step, then the rest with sigmas_2
Get sigmas as float: Just get first - last step to be able to inject noise inside a latent with noise injection nodes.
Graph sigmas: make a graph of the sigmas.
Aligned scheduler: selects the steps from align your steps.
Differences:
- force_sigma_min: off / 10 steps: gives the same values as Comfy's implementation, which matches the aligned steps of the simple scheduler.
- force_sigma_min: on / 11 steps: the added step corresponds to the minimum sigmas of the model.
- The main difference is that it takes into account the min/max sigmas of the model rather than those from the linked page. This might be beneficial with COSXL models for example.
Manual scheduler: uses eval() to create a custom schedule. The math module is fully imported. Available variables are:
- sigmin: sigma min
- sigmax: sigma max
- phi
- pi comes from math
- x equals 1 for the first step and 0 for the last step.
- y equals 0 for the first step and 1 for the last step.
- s or steps: total amount of steps.
- j from 0 to total steps -1.
- f gives a normalized from 1 to 0 curve based on a reversed Fibonacci sequence
And this one makes the max sigma proportional to the amount of steps, it is pretty good with dpmpp2m:
max([x**phi*s/phi,sigmin])
This one works nicely with lms, euler and dpmpp2m NOW ALSO WITH dpmpp2m_sde if you toggle the sgm button:
x**((x+1)*phi)*sigmax+y**((x+1)*phi)*sigmin
Here is how the graphs look like:
The Golden Scheduler: Uses phi as the exponent. Hence the name 😊. The formula is pretty simple:
(1-x/(steps-1))**phi*sigmax+(x/(steps-1))**phi*sigmin for x in range(steps)
Where x it the iteration variable for the steps.
Or if you want to use it in the manual node:
x**phi*sigmax+y**phi*sigmin
It works pretty well with dpmpp2m, euler and lms!
The karras formula can be written like this:
(sigmax ** (1 / 7) + y * (sigmin ** (1 / 7) - sigmax ** (1 / 7))) ** 7
Using tau:
(sigmax ** (1 / tau) + y * (sigmin ** (1 / tau) - sigmax ** (1 / tau))) ** tau
With a formula based on the fibonacci sequence:
(sigmax-sigmin)*f**(1/2)+sigmin
More steps means a steeper curve.
Example with this formula:
Here is a comparison, the golden scheduler, using my model Iris Lux :
Karras:
Here is a mix using dpmpp3m_sde with 50% exponential, 25% simple and 25% sgm uniform: