This model has similarities with the internal lora created by JollyLM, except this one is focused on creating both the pred and the prey in the same picture.The concept is well understood, but an idea often requires several generations if you want to get the internal part right. It is quite iffy, given that this part is not primarily affected by the prompts. It's far from perfect, but the concept is anatomically very challenging for the ai. Inpainting just the internal part with a denoising strength around 0.6 might do the trick if you really want it though, I never tried it.The model is quite compatible with diverse checkpoints and aesthetics, although it still often pushes its own on the generation. Like my other models, if a specific idea or composition doesn't seem to render well, it might just be something that the lora won't be able to produce with quality.The trigger word is "internal lora". Long prompts, especially if some elements have emphasis (1.2, 1.3...) require to augment the strenght of this trigger. For this reason, I prefered in most of my examples to keep the prompting relatively light. You might have more success than me with this though.I'm not sure if input "1 character" or "2characters" make a difference, but the lora is able to generate multiple preys type of picture, even multiple preds, although this is iffy and I'm not always sure of how those two are distinguished.When prompting, it will mostly affect the pred (the main character), and the prey will often be naked and looking like struggling or scared, because that's how the concept was understood. I feel like putting some prompts like "wide open eyes crying" helps with the face of the prey, and does not affect the pred. For that reason though, it's hard to adjust how the prey looks like.Using latent couple might allow for more control on that aspect, but I'm not savvy in it and haven't tried it.I've tried a little to generate some monochromatic pictures, to obtain a manga-like style. I feel like it works well.Eular a at 20 steps works generally well for most models. I use R-ESRGAN 4x+ as my upscaler at denoising strength 0.2, orelse it absolutely distorts the picture.I put lots of negative prompts, but they aren't required. Easynegative probably is, I don't know about the others. I put them preemptively.Although I created it, I'm really no expert at using this lora, so feel free to post your best results so I can learn from you too.
