Ep 5: KSampler settings

how the ksampler works the ksampler takes a blank canvas filled with random noise (think tv static) and refines it step by step into a finished image each pass makes the image clearer shapes form, edges sharpen, colors develop, textures appear every setting on the ksampler controls some part of this process seed the seed is a number that determines the starting noise pattern same seed + same settings + same prompt = same image every time fixed keeps the same seed useful when you're tweaking other settings and want to see what each change does in isolation randomize new seed each run what you want when exploring increment/decrement moves up or down by one each run steps how many refinement passes the ksampler makes more steps = more detail, but diminishing returns going too high can overcook the image the right number depends on your model some turbo models work in 4 8 steps others need 20 30 check your model's documentation cfg (classifier free guidance) how closely the ai follows your prompt low values more creative freedom results may drift from your description but can feel more natural high values sticks tight to your prompt, but push too high and you get blown out colors and artifacts cfg at 1 your negative prompt is effectively off set cfg to whatever the model creator recommends and leave it this isn't the setting to experiment with first sampler and scheduler these work together the sampler is the algorithm (euler, dpm++, etc ) the scheduler controls the pace of noise removal at each step euler is the most common default use the model's recommended settings denoise controls how much the ai is allowed to change the starting image this matters most in image to image workflows (covered in episode 8) value what happens 1 0 ignores input entirely, generates from scratch 0 5 0 7 keeps the general structure, takes creative liberties 0 2 0 4 gentle changes, preserves colors and composition 0 0 changes nothing for text to image with an empty latent image, leave denoise at 1 faq what seed should i use? there's no "good" or "bad" seed it just determines the starting noise if you find an image you like, note the seed so you can reproduce it how many steps do i actually need? depends on the model start with the model's recommended value going way beyond that wastes time without improving quality what does cfg 1 mean? maximum creative freedom, minimum prompt adherence your negative prompt stops working at cfg 1