from Hacker News

Control your AI images with these must-know parameters

by andrewon on 10/28/22, 6:21 PM with 3 comments

• by orbital-decay on 10/28/22, 11:50 PM

  Using euler_a (or any other ancestral sampler) with SD is bad advice and makes all examples flawed. It's just a bad default in automatic1111's web UI. Compared to the normal euler it just mutates the output every step, so you're just making it less predictable and harder to tweak. In fact, most of these samplers are shown in the UI just because they were included in the third-party samplers library, they don't necessarily make sense for the better output. In particular, heun solves the stochastic ODE with Runge-Kutta, which makes it roughly twice as slow but it converges in twice as few steps as euler; this just leaves you with less granular controls. In fact, you can use just euler for everything. Or try DPM2_Karras, it's usually a bit faster, or DDIM.

  Either way, it's just txt2img and it's not very interesting by itself; the real power of those models is in style transfer and the ability of being trained on custom stuff, in a variety of ways.

• by brnaftr361 on 10/28/22, 8:58 PM

  Also I've found that using various aspect ratios does not add artifacts, and that it does have implicit data e.g. most portraits are taken in... Portrait, and so if one is intent on fabricating a portrait, such an aspect ratio does, in my experience, tend to generate more favorable outputs.

  There are several other features that aren't listed here, such as tokenizing, weighting, AND, img2img, feedback... Textual inversion, and aesthetic gradients...

  Another way I've been fairly successful in, is using 1- or 2-step outputs to find a good basis for composition as you can generate a lot of outputs quickly and take the best of the bunch without waiting eons for your 16x16's to diffuse.