Most of the time it is quite effective, probably.When the prompt is very long, Lora weight 1 should be used, and when the prompt is relatively short, weight 0.7 is generally sufficient.It will affect the character actions, hair color, etc. of the model.The trigger word designed for Lora during training is 'cumdump', but it seems that using 'bukkake' and 'excessive cum' would be better. I have another Lora model designed for chenkinNoobXL_v01.safetensors, but it has almost no effect on the Illustrious series model. Is anyone interested?noob 1.1 version is a lora, not a lycoris, I don't know how to make individual changes. Please use lower CFG scale. Lower CFG(like 4-5) results in more cum, maybe. ^_^ Sorry, it is not stable enough. Due to my lack of ability, I tried for a while but did not achieve good results in this concept. Maybe you could try it with some artist tags. You can add "projectile cum" in negative prompt to avoid cum in background somehow (not always). Warning: Sometimes generate really bad anatomy picturesIn the recent version of training, I used too many images drawn by Yamataka, which caused a change in the art style. However, the model generates more liquid and is relatively more stable......................................You could edit the weight of "excessive cum" to generate more cum. (longer prompt needs higher weight like (excessive cum:1.4))I suggest placing trigger words in a more forward position.....................................The pony1.1 version of the model is not an upgrade to the previous version, but rather a test with relatively few training steps. However, its performance is still good for, and its impact on the art style seems to be relatively small, and it can generate enough cum with longer prompt (the previous version does not). (<s>I don't know why</s>) In short, I uploaded that file and it was expected to be trained for 16 epochs, this is the file from the first epoch, and I found through actual testing that it works the best (which is indeed quite strange). If you want the files of other epochs, you can DM me. I trained until the 6th epoch and then terminated the trainingPony 1.0001: I personally prefer a more 2D art style, and this model has a greater impact on the art style compared to the previous one (because I trained more steps this time)The larger file (. safesensors): pony_excessive5-000016 The smaller file (. zip): pony_excessive5-000010 This version only changed the lr=0.5 and d_coef=2.0 of version 1.0 to lr=1.0 and d_coef=1.0, and conducted a complete 16 epochs training (the 1.0 version of pony_excessive1-000010 should have been trained for 16 epochs, but I terminated it at the 10th epochs because the preview image generated at that time was a bit unsightly, and I mistakenly thought the training had failed at that time).....................................You may have noticed my naming of Lora files, especially for the Lora of the Animagine model named Myfetish. So the dataset not only contains images of excessive cum, but also images of wounds (bruise/injury/ryona/guro). If you want to try those contents, please try "excessive+ bruise/injury...". Some of the versions also support "pee, pee on body" etc.Tips for early versions:v1.1Tips: v1.0 and v1.1 could be used together, example: &lt;lora:an31_excessive_cum:0.3&gt;, &lt;lora:excessive_cum_v11:0.9&gt;. Reduce cum on background: please use [:(excessive cum:1.1):0.3] instead of (excessive cum). I used the Lora I uploaded last time and some other Lora(s) that were previously trained to do some random merges. In short, the new Lora seems to have reduced its impact on the art style and will not produce bad images when using "dirty" and "cum" in the positive prompt words at the same time.Lower CFG Scale ≈ more cum (but worse hands) "excessive bruise" "excessive injury" "excessive scar" are still available. Problems: "gokkun" doesn't generate cum drinking. It generates more cum on body.v1.1 is bad at drawing pee on body, but you could try it. If you are using Animagine XL V3.1, you could try artist names. I didn't trained with artists names in tags, but it doesn't work bad if I add some artist names in prompt. The blue hair one used "mana \(remana\)" and "liduke".v1.0 Maybe not excessive enough. The main purpose of this model is not actually to generate massive cum..Tips: Animagine V3.1 itself can generate cum bath well at CFG scale 3.5 without any lora. Please don't use "dirty" in prompt when you are generating cum using this lora. The breasts become too large if you don't add "huge breasts" in neg prompt. Problems: Often generate dirty cum(instead of "white" cum) lying: often generate bad images.

Excessive Cum (or Scar), Bukkake - AnXL,Pony,Noob

It should be the last version. <s>After spending some money trying to train a Lora that satisfies me, I only have 200 RMB (27.6 USD) left for living expenses this month, QAQ.</s>This version only changed the lr=0.5 and d_coef=2.0 of version 1.0 to lr=1.0 and d_coef=1.0, and conducted a complete 16 epochs training (the 1.0 version of pony_excessive1-000010 should have been trained for 16 epochs, but I terminated it at the 10th epochs because the preview image generated at that time was a bit unsightly, and I mistakenly thought the training had failed)There are two files, the larger file (. safesensors) is a 16 epoch file, and the smaller file (. zip) is a compressed file of the newly trained file trained to 10 epochs

Files with SHA256: 8207f8309d4c54f396d300ac1a999b02fdcf8738dab2d0a48bd9ed1388678643