Plushtrious Split Conditioning with Sliders and ControlNet

Nothing exciting here. Just sharing the way I like to do the pony. I remember being blown away the first time I started playing with chaining ksamplers. It's a really powerful concept. The best part of this is that you can mix models. I stumbled on pony + illustrious after playing around with combos, before discovering that tons of people have been doing this forever, of course. Anyway it's great. If you like the style of pony but not the fine details. But of course you can mix and match whatever models you like.

As posted it is set up with a pony model controlling ksamplers 1 and 2, and an illustrious realism model controlling 3 and FaceDetailer. Each model has its own sampler/sched. nodes. If you enable ControlNet you can utilize depth maps, poses, or any other controlnet model. I previously posted a depth map workflow for saving favorite poses. It saves a hell of a lot of time.

The LoRA section is a bit convoluted, this is the way I use it so it's not really set up for clarity of flow. As this isn't exactly original I don't see any point in cleaning it up. It's just a bunch of stacks controlled by sliders. Two stacks for female and one male stack, each has a boolean switch. You'll have to patch in your own filenames of course. Some have their original names but at a certain point I realized how important it is to subfolder and rename them with useful information along with weight range for slider types. It's definitely worth your time to set up slider stacks like this and save them as templates. The stacks themselves are minimized, as you don't need to see them when you've got sliders.

The conditioning section might also seem convoluted, but all of it is concatenated in the end. I got sick of having to hunt around like a pirate for buried tokens, so I broke up the positive into sections: General Character, Environment/Style, Facial Features and one for the refining stages that only focuses on the face. If you spot some choices that may cause dilution, good eye. There are infinite ways to do this, most are probably better than mine. But adding switches to the conditioning always causes problems for me. So I just bypass or renoodle if something isn't coming through.

There is a kj node at the beginning that lets you select preset sizes for your latent. In order for it to work you need to find the node's folder add your own text list, I think it comes with a template but it's got '_example' appended to the filename, rename it 'custom_dimensions.json'

Because the prompts are simply concatenated, you can bypass many of the text boxes. Two will suffice I think. Different effects can be achieved by meddling with your ordering, the same as changing the order of prompts in a single box. Diffusion can be very stubborn in this configuration. Great if you want consistency, annoying if you want extreme variation. I've bypassed some inputs in a few of the example generations as examples of how to get around diluted tokens.

That's it. Keep in mind this is geared towards specific character features, not intricate environmental stuff. Skin, facial features, fine details. Stuff like that. Maybe you'll find it helpful.