Anything can NSFW these days, but as the model is heavily influenced by Animagine, it will be harder to generate images of men interacting with women while they are fully/partially visible without extensive prompting. POV and basic NSFW are easily achievable in base Animagine as well, so this model keeps up and (hopefully) enhances that experience. Ohh and make sure to use the newest, Gamma version which I have just uploaded

Thank you. By the way, as this is an Animagine-based model, you don't really have to use a lora for Love Live girls. It already knows them. You can call them out in your prompt by their danbooru tags, like this: "yuki setsuna \(love live!\), love live!, love live! nijigasaki high school idol club"

This is pretty much the entire reason for the model, to not require LoRA for popular characters, just like on Animagine, but to embed the better anatomy of Pony.

Here's an example I made in a rush testing out Yuki Setsuna

@Ainme_Diffusion I'm glad you can combine loras like that, thank you for the feedback! I just intended to make you aware as a lot of people have been using only pony-based models and they are not aware of the extensive character/artist knowledge of other model branches. Great work with the style!

Thanks. Animagine itself might be better at making cute stuff though. At least it has proper backgrounds, unlike my model.

Indeed, for the gamma version the score tags don't do anything. This can be easily determined by adding 'text' to your prompt as it will start writing 'score' everywhere on the image as it doesn't understand the tag.

The Second image has my entire workflow. You need to use comfyui to replicate it. Just click on the little (i) button, then on the copy sign where it says Workflow: 117 nodes (copy sign).

Without Comfyui, the token weighting is entirely different so there is no point in trying to replicate it on A1111/Forge

If I can find the original Op image, i'll share it with metadata

Here is the original image with metadata. Again, it's Comfy.

https://files.catbox.moe/b49ilc.png

score_9, score_8_up, score_7_up, source_anime, takane lui, hololive, takane lui \(1st costume\), hololive, absurdres, highres, 1girl, barefoot, bed sheet, black choker, black gloves, black shorts, breast zipper, breasts, choker, closed mouth, collared shirt, feet, foot focus, foreshortening, gloves, greek toe, hair wings, hand up, head wings, indoors, knees up, large breasts, legs, legs together, long sleeves, medium hair, necktie, one eye closed, partially fingerless gloves, pink background, pink hair, red necktie, shadow, shirt, short shorts, shorts, sidelocks, soles, solo, thigh strap, thighs, toenails, toes, tongue, tongue out, virtual youtuber, white shirt, wings, xo \(xo17800108\), expressive, award winning, 8k, absurdres

@User37 It loads just fine if loaded in comfy and you install all of the custom nodes (as the prompt is written in 8 text boxes provided by custom nodes which are then merged together in a specific order)

It ends up looking like this: https://files.catbox.moe/89pd51.png

The end result after merging the text boxes is:

Positive:

score_9, score_8_up, score_7_up, source_anime, takane lui, hololive, takane lui \(1st costume\), hololive, absurdres, highres, 1girl, barefoot, bed sheet, black choker, black gloves, black shorts, breast zipper, breasts, choker, closed mouth, collared shirt, feet, foot focus, foreshortening, gloves, greek toe, hair wings, hand up, head wings, indoors, knees up, large breasts, legs, legs together, long sleeves, medium hair, necktie, one eye closed, partially fingerless gloves, pink background, pink hair, red necktie, shadow, shirt, short shorts, shorts, sidelocks, soles, solo, thigh strap, thighs, toenails, toes, tongue, tongue out, virtual youtuber, white shirt, wings, xo \(xo17800108\), expressive, award winning,

8k, absurdres,

Negative:

text, japanese text, error, blur, blurry background, multiple views, source_pony, source_furry, source_cartoon, ,

This website will also show you the metadata but they delete results after 24 hours (you can load the image in yourself if they deleted it): https://www.metadata2go.com/result#j=6d22d32d-c1c6-467a-ab01-59cab75c8af0

And again, replication is impossible outside of comfy because A1111/Forge use a different normalization of tokens.

@User37 You don't need everything to get the image. You can pretty much stop the workflow at the point where the second ksampler outputs an image. I think the group is called "2nd pass" or "hi-rez pass". Everything afterwards that point was not used to make the images and just exists in my workflow, so no Adetailer and no Ultimate Upscale would be required.

It's is not being used in this merge. I have it left over in the workflow because I wanted to play around with it but my RAM couldn't handle everything. If I upgrade any time soon and try it out, I'll post another comment. For now the only explanation is the one available on https://github.com/54rt1n/ComfyUI-DareMerge?tab=readme-ov-file#normalization

@TheKite You can check out my new model merge released earlier today, LeekAI https://civitai.com/models/564727?modelVersionId=629367

@TheKite That one didn't use DARE merging. It used Block merging (MBW) instead. It does have a tiny insignificant fraction of this model in it though, inherited from Trifusion which used Ponymagine as base.

Try artist names as styles similar to how animagine does it, by starting your prompt with 'by xxxx', however if you want styles, I would recommend my other model, LeekAI, which is far better at changing its style on the fly. The intention for Ponymagine was just to have better compatibility between Pony and Animagine LoRAs.