Welcome to Eros Ultima
--
Why I Make This Model
This is my attempt to combine Pony base model with SDXL base model and since i like to train and use my own lora, But I find it difficult to get a pony model that can translate my Lora correctly. Generally, XL model is better for that, but most of the XL model has bad NSFW / Anime prompt interpretation, maybe due to their safety filter in place, pony and anime model on the other hand, can generate a lot of NSFW that is not available in a realistic model.
--
When to use This Model
Good compability with SDXL / Pony Lora, Eros Ultima has a very low bias, so Lora will be represented correctly 90% of the time.
Use it mainly for Erotica, this model is merged using pony anime model that has a lot of NSFW fidelity, but with SDXL realistic model.
Very NSFW... use at your own risk!
Hybrid Model
This model is better at interpreting anime / conceptual background, but at the expense of 20-30% of realism out of the character such as the skin.
Realistic Model
This model is still better at interpreting anime prompt, and translating it to realistic model (similar to other pony realism model, but with low model bias), however at the expense of not lesser interpretation on the fantasy / conceptual style or background.
--
When problem arise...
If the image turns out "burned" try lowering CFG scale or your lora weight, everyone trains their lora differently, your lora may be overtrained according to this model.
Whenever a body parts appear mangled, it may be also because the same issue as #1. Try to use a negative lora, or adetailer first, before lowering CFG scale or lora weight
Description
FAQ
Comments (5)
Underrepresented model! I really like this! A little rough with some things but composition and prompt adherence are great. This is now my go-to for first-pass realistic generations. I found that refining or img2img'ing with Lustify gives some fantastic results. Thank you!
This is the best model for my 10c.
Tends to do hands better than some high-rated realistic models so it saves a lot of render time. Lighting tends to be a bit flat by default but I like that because there are fewer shadows to obscure "important details".
This is my preferred model, and it's twin the "realistic" variant. Thanks for your great work!
I found it's a lot more faithful to the tags that other popular choices such as Everclear.
We do get some "interesting" results sometimes though, for example I asked for a cyborg wearing a "revealing swimsuit", and also mentioned that the setting should be in Canada.
She ended up with Canadian flag nipple-pasties! Maybe this is the Olympics of the future... It's a bit of a problem with pony, mentioning a country tends to produce the flag. Maybe I need "flag" in the negative prompt.
Regarding unexpected or seemingly random "mangling", I found that adding lots of "BREAK" throughout my large prompts and even in the negative prompts can help a lot with this. Any time automatic1111 would extend the prompt length, use BREAK instead. For me, it's giving much better results.