I’ve tried many sampler/scheduler combinations for Krea 2 Turbo. In my tests, the standard 8-step recipe often gives either residual noise or not enough detail. So I built a workflow that gets the most out of Krea 2 Turbo INT8 — close to RAW quality.
Description
FAQ
Comments (6)
I'm not entirely convinced. A long time ago, thanks to Reddit, I discovered that combining ER_SDE with SGM Uniform and using 12 steps gives the same, if not better, results than what you're showing.
You are right when it comes to the base Turbo model plus a maximum of one LoRA. However, when dealing with a model that has multiple LoRAs built in, the result is a conflict that manifests as noise. That is precisely what my workflow addresses.
My process needs some tweaking to make the image brighter; I’ll work on that in the future.
Thanks. With your tip, I’ll be able to create a more effective WF. Now I’ll select the parameters to remove unwanted artifact noise and add useful noise.
@Viennar no problem B-). If we're already chatting, I'd like to ask you about your EXERCITUS model. I'm really fascinated by generating images with an amateur, old-digital-camera look, where the women in the photos look like ordinary, average people rather than models. Does your model handle this kind of style well?
@Lombard99 The model has been trained on dozens of styles. In the example on the homepage, I demonstrated "pure" styles that can be combined by mixing prompts.
@Lombard99 https://civitai.red/posts/31275606 a few examples