KSampler Overview

Here we are going to have a brief overview in non techincal terms of the KSampler and what each option does. 

This is what the KSampler looks like:


The KSampler is what condenses the maths behind the scenes into a readable format for the VAE decode (or whatever the next step is). There are quite a few options here, and no real explanations, so here goes.

Seed:                                This is the random number the generator uses to create the images.

Control_after_generate:    You can fix the seed so it doesnt move, increment or decrease the number by 1 after every generation, or randomize so it will be completly different every time. 

I usually lock the seed if I am testing a model, or after I may have found a particular look that I like and dont want the seed changing and losing the image style. The changes are subtle, but they are there.

Steps:                             Generation works in iterations. The steps on the sampler are how many iterations it will do to a particular image. The less steps, the more noise will come through in the image (think static). and the more steps will iterate more, giving more detail and more chance of things working out. I start with 20 normally, and use 30-50 steps when upscaling the images. The more steps, the longer the image will take to generate.

CFG:                                This is a number that tells the generator how important the conditioning prompts are. A cfg of 1 give the generator free reign to do what it wants, leading to some crazy images. The higher the number, the more constrained the to the prompt the generator will be. I usually use between 7 and 12, depending on where the model is taking the images. The higher the cfg however, the more saturated the image colours will become, eventually distorting the image.

Sampler name:                 This is the type of sampler used. This to me seems very opinionated as to which one to use, I use euler and dpmpp_2m as I do a lot of human photography style images. This is one to play around with your self, lock the seed and see what changes the sampler makes.

Scheduler:                    This works with the sampler. I use Karras a I feel this creates the best images for me. Again, test your self between the options to see what is best for you.

Denoise:                      The amount the generator will change from the original noise mesh that is created. I always have it on 1 for generation, and use 0.3 - 0.6 when upscaling. This is a interesting option that can do many different things. A low denoise generally doesnt work for a main prompt I have found as it does not distill the image down enough. A high denoise on a Img2Img will completly change the original image however, so lower is better here.

Comments

Post a Comment

Popular posts from this blog

Basic Nodes and using Comfy UI

Image
 What you see when you start Comfy UI are the basic building blocks of the interface. There are more complicated, more diverse, more niche case options, but we will come to those later. Lets take a look at what we have now. This is the basic screen that you will see when you click the load default button.     What does all this mean? Lets start from the left hand side:   Load Checkpoint:   This is the node that loads the Checkpoint, or model, that the ComfyUI engine will use to genereate images.  The three coloured dots on the left hand side are the outputs that this node uses. Model goes to KSampler, Clip goes in to the conditioning prompts, and the VAE is used to decode things. The colours will all match the corresponding inputs. Click the right arrow will pull up a menu with all the Checkpoints that you have installed. Clip Text Encode (Prompt): These nodes are the basic building blocks of ComfyUI. You tell the engine here what you want, in the top box...
Read more

Output Images and where to find them

All the output images are that are run through the save image node are saved in the output folder. Where can this be found? In the same directory as the Model folder we found in the last step! It is towards the bottom of the folder, aptly named "Output". Double click and all the images that have been saved will be found here. They are saved as .png files, and can be edited in photo image software. Remember to resave them back as .png if you want to put them back into Comfy UI. And one really cool thing, if you like the look of a output image but cannot remember how you generated it, just drag the image back into the Comfy UI web interface, and it will reload all the nodes, prompts and settings that created the image! Have fun!
Read more