SD1.5-LibertineMix

WARNING: HERE BE DRAGONS!

This model is made though merging with ComfyUI as well as using a custom CLIP. Compatibility with existing architecture is maintained. However, potential performance between actual model and reality may be different. User discretion is advised.

"A libertine is a person questioning and challenging most moral principles, such as responsibility or sexual restraints, and will often declare these traits as unnecessary, undesirable or evil. A libertine is especially someone who ignores or even spurns accepted morals and forms of behaviour observed by the larger society." - Wikipedia, Libertine

What is this model?

This model is a 2D anime-styled Stable Diffusion 1.5 merge using NovelAI v2 as the base in an attempt to "migrate" some of the trained knowledge in many of my merges forward into a more suitable base. The main idea of this merge is to use a combination of TIES (and DARE-TIES-adjacent methods) and block merging to have as much granular control as possible on the final merge itself. This also has a side-effect of making this model more robust in higher resolution ranges up to 1024x1024 and beyond (though still not foolproof for SD1.5).

Why is this model a thing?

Every once in a while, I do large merges - merges that requires 10+ models at once, just because I got a bit of an ego and wanted to see what would come from them. Many of these attempts (SoleraV6 and V8/Pandora's Box v5/ColorSplash-v0.1.4/OpenSolera-a8 and a9) never saw the light of day because of how bad they were compared to the amount of effort I put into them. This might be one of the few exceptions to that. Now, with NovelAI v2 being around for a good while as well as a bunch of good models I've already used and merged with, I thought "hmm, why not try and 'migrate' to NovelAI v2 base?" This is that attempt at it, with most of the methodology for this model coming from my own Pandora's Box series and a bunch of experimentation beyond that.

How do I use this model?

Prompting

Similar to basically most, if not all anime models in SD1.x and SDXL, this one uses booru tags. Despite that, there are a few additional things to keep in mind: unlike many of my other models, this one is far more temperamental at higher token count, and is really hard to steer from time to time. You can get beautiful results from it, but actually taming this has been a challenge on my side.

NSFW prompts do not work as well as I wanted to here, as it is much harder to get decent results compared to its base, but it still works. Same with many other pieces of popular domain knowledge (including Hololive and some monster girls, except many undertrained ones in AIOMonsterGirls)

Positive prompts: Doesn't really matter, but I default to:

high quality, absurdres

However, your standard quality prompt will function just as well.

Negative prompts:

worst quality, low quality, lowres, interlocked fingers, zombie, text, watermark, signature, loli, shota, chibi, <insert whatever else you dislike or don't want here>

(Yes, this is partially a reused one from MeinaMix. It works. Add and remove based on your requirements. This model doesn't actually do a lot of signatures, this is mostly a safety add)

Other tags: YMMV for a lot of them, but most should work with some fiddling with weights.

Resolution

As far as I can tell, this model supports resolution up to 1280x1280 with relatively manageable distortions and up to 1024x1024 with minimal to no distortions. Despite that, recommended resolutions would be around the 768x1152 and similar resolutions.

At higher resolutions, this model can handle 3:2, 4:3, and 16:9 aspect ratios for both portrait and landscape. However, for best results, 3:2 is recommended.

Parameters

• Step count: 20-50
• Sampler/Scheduler: Anything works. I'm more partial towards Euler Ancestral CFG++/Beta and UniPC/DDIM Uniform.
• CFG: 6-11
• CLIP Skip: 2 when possible
• Secondary hi-res pass is optional but recommended.
• Platform: Tested against ComfyUI 0.18.1 with --deterministic. Results with A1111-like platforms may differ.

License

This model attempts to follow and respect the licenses of its merges to the best of its abilities, including the following:

• CreativeML Open RAIL-M
• Fair AI Public License 1.0
• Creative Commons BY-NC-SA 4.0 (inherited from NovelAI v2)
• SD-Silicon's Terms of Use

If the author of any of the base model used in this merge has a complaint about this model, its licenses, or more, please send me an e-mail over at NexusLightvale(at)protonmail.com.

This is a non-commercial model, meant for personal and research use. You cannot sell this model, its merges, or its direct image generations. In order to use this model, you don't need any attributions back to me, but it is appreciated regardless. I am not responsible for what you do with this model, that is on you as the user. This model is provided as is, with no warranty attached.