Style of the artist <a target="_blank" rel="ugc" href="https://rule34.xxx/index.php?page=post&amp;s=list&amp;tags=muchin+&amp;pid=0">muchin</a> <a target="_blank" rel="ugc" href="https://tensor.art/models/721889427548400521">Tensor.art</a> | <a target="_blank" rel="ugc" href="https://yodayo.com/models/3d65ef06-98b4-440e-846a-3cd5e5632b47/">Yodayo</a> | <a target="_blank" rel="ugc" href="https://pixai.art/model/1741478164697730239?utm_source=copy_web">Pixai.art</a> | <a target="_blank" rel="ugc" href="http://Seaart.ai">Seaart.ai</a>  This is a quick post I tested a version, but I don't have much time to describe it. Thank you for your activity. The style works great on NAI is really similar. The new pony version is not factually better, but probably at least as good. I haven't had much time to compare either.UPD. and so was able to buy myself some time for more testing NAI version seems not always stable, but below I seem to have indicated the best option, the style is kind of closer to the original, but at the same time when how. The new ntr mix version seems to affect the style less with the quality tags. But still I can't say it's better to use them or remove them all. I know how to refine these, but I don't have the time, so there is what there is. I personally like the facial expression of the model's faces, especially in nsfw Illustrious/NAI XLI should add version decoding, 1.7 is trained on NAI 1.0, 1.8 on nai 1.1 that seems to be the end of their differences. Just because a version is almost 2.0 it just means I label each version as the next, it doesn't mean there are global changes. I can't say that 1.8 is better, it's more a matter of compatibility, I would say both versions have the basic problems of the first versions of pony, (like 0.94) blurry background, artifacts, fuzziness, but also something of their own, for example for some reason sometimes the eyes become fuzzy, I had to increase inpaint noise strength of adetailer, I don't have time to find out what exactly the problem is, but maybe it's related to the specific model I'm generating on now.Trigger word: 0tx1 Recommended weight: 0.7-1.3 (1.2)Recommend cfg scale: 3-5 (4)at least for <a target="_blank" rel="ugc" href="https://civarchive.com/models/926443?modelVersionId=1061268">this noobai xl model.</a>UPD. For NTR MIX XII Recommended weight: 0.7-1.3 (0.8) for original style 1.1Recommend cfg scale: 4-5.8 (5)v1.4 PDXLThe version is likely to be intentionally a bit more overcooked than the others, the version is a bit messier, but the style is supposed to be rendered more accurately. The version is aimed at standard color generation, the dataset for example has been reduced a bit. Most likely the version is not fundamentally different from the old ones, they were already good, you can consider this experimental and try.upd. Having done a little testing I would generally say that this version is probably the best (among the pony versions), so try it out.Recommended weight: 0.7-1.1 (0.9-1)Trigger word: 0tx1 For all PDXL versions: Minimum recommended steps: 28 In general sfw sense style has little to no effect on anything, mainly on the eyes, on skin color, as well as in the case of for example ponyFaetality 3d takes more 2d stylistics, on the environment stylistic effect, especially not. Works with all source, if it's not a human sometimes a little lora weight reduction is recommended. Preferably not less than 0.55. Recommended to use adetailer, hiresfix. All images have a prompt you can look at and choose what you are interested in. v1.1 Animagine XL Differences from v1:<ul><li>in the dataset a little less not original and a little more original style, as well as a little more censorship, in fact nothing has changed it's just a version for sdxl.</li></ul> Trigger word: fmuc Recommended weight: 0.9-1.2 (1) for the original style maybe 1.1Version has much more stability issues especially in terms of nsfw. Training on Animagine XL v3.1 v1.0 PDXL Differences from 0.94:<ul><li>Almost all the censorship is removed, as is the comic book style.</li><li>The smegma tag would theoretically work a little more consistently and better.</li></ul> Trigger word: mainmuchinponyv1 or 0tx1 You can also try: fmuc for 2d Explanation: I was told that the style in v1 is not original enough, well the request is fulfilled I made a broken version, which is nevertheless closer to the original than v1. Exactly because it's broken, that's what the trigger words are here: 0tx1, mainmuchinponyv1, supmuchinponyv1, lowsupmuchinponyv1. And so, it's kind of like pony tags, the word kind of hits the right folder, BUT! It's not done intentionally, so it's basically meaningless, 0tx1 is the connecting word for all folders. I think 0xt1 is less oversaturated and over-lit, but only a bit. You can try to get the old muchin style out of lowsupmuchinponyv1, but I can only wish you luck, it's not designed for that and it's also very similar to another trigger word, you can find comparisons between them the point is that mainmuchinponyv1 is already the largest folder and also the most trained given the increased number of steps, so in most cases it's enough, but maybe in narrow scenarios it can only get in the way. 0tx1 has roughly the same effect, I haven't tested much you can compare it yourself, maybe it gives more dataset coverage, but that's not certain. THE VERSION TAKEN IS INTENTIONALLY NOT OVERCOOKED. So, you are bound to use at least some trigger word unlike v0.94. If you want the style to affect less, it's better to just use the v1 version instead of not activating the trigger word. Nevertheless, there are different options in the example images. And yes, for some reason the word fmuc (maybe because of the muc in it) affects the style a bit, making it more 2d, you can see the comparison for yourself, I don't have time to experiment, but given the size of the trigger words maybe by letter you will find a better performance or combinations of trigger words will give you an unusual effect, I don't know. The word recommendations for v1 are generally still valid, although some of them make less sense. This version has the same problems as 0.94 with skin color, background blurring, overlighting, although of course for those who are used to 0.94 is a plus, yes, I believe that you will lose nothing but censorship and characters if you start using it instead of 0.94, but instead of v1 only if you really lack the original style. It is also possible that the trigger word is too concentrated on one type of character, that is, if you do not specify anything in the girl, she will have said the same hair color almost always. Recommended weight for original style: 1-1.1 (1)If you are going to use this style as an additional style, I recommend looking at the more flexible version v1 Recommended negative prompt words if you want more default generations: bar censor, mosaic censoring, tan, oversaturated, yellow skin, blurry background, greyscale, monochrome, 3d Non-obvious points: If you use xyz in local web ui, there may be problems with prompt s/r, so put in any word like fmuc and only then say 0tx1, otherwise it will find that text at the lora file and break the lora work. V1 PDXLDifferences from 0.94:<ul><li>Almost all the censorship is removed, as is the comic book style.</li><li>The smegma tag would theoretically work a little more consistently and better.</li><li>More flexibility including weight.</li><li>The background is more often unblurred now.</li></ul>Trigger word: fmucRecommended weight: 0.6-1 (0.85) for the original style 0.9-1 Prompt recommendations: Recommended words for muchin original style: score_7, anime screencap, flat colorsFor muchin style sex: 2girls sex, futanari fuck girl, futa with female, cum, penis shaped stomach bulgeAlso: brown penis, dark veiny testiclesFor face style: chestnut mouth, serafuku, blushAlso: sanpaku, upturned eyesThese words may also be useful: cum overflow, muscular, motion lines, motion blur, trembling, x-rayRecommended negative prompt if you want less yellow skin: tan, oversaturated, yellow skin Possible problems/Non-obvious points: By default, the more lora weight, the more muscular the characters will become. Without negative tags greyscale and monochrome lora likes to produce such images. The heavier the lora weight, the closer to the original style, which means the skin color will yellow out. Maybe in special cases in the background will be divorces on the texture of objects, you can in this case reduce the weight of lora or specify for example the depth of field, although it is difficult to get such a thing. Thanks for your activity, it was the main motivation to improve lora. I left the trembling effect, so it can be annoying, but removed all other visual comic effects and censorship. Also, I noticed how many people liked smegma, so I upped it priority a bit. This is the last update for a while, though lora may still be tweaked a bit, including a special facial expression option. But overall lora is ready and stable, the dataset has been diluted a bit, so the original style is a bit lost, but it's still very close. My tests have shown that there is no point in a 1.0 version, as I sometimes do. Also theoretically the style knows all the right characters, and also theoretically the variation of men has gotten a bit bigger, maybe variants of extreme concepts have become easier to implement, but I haven't tested. v0.94 PDXL The description below is divided into 2 variants, a generic short sfw variant, and below a detailed nsfw variant. General style recommendations: Trigger word: fmuc (upd. doesn't seem to be working, details below.) Recommended weight: 0.7-0.9 (0.75) for the original style with artifacts up to 1.2 Works fine with most pony checkpoints, SDXL 1.0 options not tested. Trained on Pony diffusion v6 checkpoint. Detailed style guidelines: From a nsfw perspective style affects many different aspects, I'll start with the positives. The genital style is similar to the original. The faces also have more emotion, in terms of ahegao. Also the sex scenes are accompanied by visual effects. More specifically: The pleasure effect on the face is enhanced, I recommend trying ahegao, fucked silly, mind break. Although the effect will also be just like that. Tense lustful face excellent fit. Recommended words for muchin style sex: 2girls sex, futanari fuck girl, futa with female, cum, penis shaped stomach bulge This words may also be useful: cum overflow, muscular, motion lines, motion blur, sound effects  Use appropriate positive and negative words, here is my version of negative prompt: score_4, watermark, signature, greyscale, monochrome, (bar censor, censored pussy, comic, blurry background, pointless censoring, censored, mosaic censoring, speech bubble, black lines:1.1), convenient censoring, heart censor, oversaturated, english text, username, 3d Recommended negative prompt if you want only less censorship: (bar censor:1.2), (censored pussy, comic, blurry background, pointless censoring, censored, mosaic censoring, speech bubble, black lines:1.1), convenient censoring, heart censor Oops it seems due to a bug in the training of sdxl models, txt files are not accepted by default so for version 0.94 you can try to use the file name as a trigger word. UPD. 2, I apologize and so actually the bug made 3 different trigger words since I used 3 folders. Its look like this mainmuchinpony0, supmuchinpony0, lowsupmuchinpony0, from left to right their quality changes, I think you know the analogy with score. So to use the default you can either not do trigger word or try 94tx1, the neural network doesn't know the word fmuc and apparently it doesn't affect it in a bad way, so it will be almost equal to empty space. Possible problems/Non-obvious points: Censored, it's hard to find the right images, censorship will occur, the more often, the greater the weight of lora. The background is blurred more often, there are particles of light especially indoors, and also goes oversaturation of light. if you don't specify specific character traits, the appearance traits of characters from the dataset will often appear. By default, the more lora weight, the more muscular the characters will become. Without negative tags greyscale and monochrome lora likes to produce such images. The heavier the lora weight, the closer to the original style, which means the skin color will yellow out. Occasionally, in the case of a particularly high lora weight, something resembling a belt may appear on the penis. Would love to see your generations!

Muchin style むーちん [Pony, Illustrious/NAI, Animagine XL]

Redone more than 80% of dataset, removed censorship, improved quality, fixed bug now trigger word fmuc works correctly, the training was on adafactor, including lr sheduler.

Files with SHA256: fd4ba1c69ab2ddbd09c439bcd8dad8c2e7fd30972bb28922f2500857fd8a95cb