Update: I've trained a new version with a much better training setup and I also trimmed the dataset and cut it into 3 categories to give much more control over the size and type of bulge, plus one version that includes all 3 sets which I highly recommend mixing in with other versions to improve consistency.Version Nice: dataset trimmed for images with nicely detailed bulges that aren't ridiculously huge (still pretty big though, I might make another version for better smaller bulges)Version Absurd: selection of some of the most absurd insertions/penetrations. This one especially benefits from mixing version "All" with it, something like absurd:0.8 all:0.6 will give very consistent outputs. I tried merging them into one lora but every method I tried to do so gave vastly inferior results compared to putting both loras in prompt.Version Pectoral: images with bulges reaching up to chest. Intended to be for impossible penetrations going right through the chest instead of the stomach bulge extending up to chest level, but it's ambiguous.Tips: Don't forget to add vaginal/anal, it'll vastly improve the penetration part. Also, I recommend adding fat or pregnant in negative to improve the shape of the bulge.LORA trained on a large set of images focused on large penetration, more specifically large stomach distention. I tried going for versatility and minimal visual impact. High epochs with high weight can get very weird, I'd recommend starting with epoch 4 and 0.8 or lower weight (v7 seems fine with higher). Even then this is meant for extreme content, if you just want normal large penetration many models can do that very well already.For more absurdity, try things like absurd stomach bulge, all the way up torso, or ask for sideways viewpoint. I've added a lot of tags to training data, but it doesn't particularly react more strongly to exact tags, just try things. Seems pretty difficult to train a LORA that can react to many specific things like the shape of the stomach bulge, I tried but without much success, at most minimal changes on occasion when forcing it with 1.5+ emphasis. (things like ribbed or spiky dildos, or cone-shaped object insertion)Combines well with other LORAs, as demonstrated in the last example image. <a target="_blank" rel="ugc" href="https://civarchive.com/models/7920/extreme-insertion-preview-in-description">Extreme insertion</a> can improve the penetration part, <a target="_blank" rel="ugc" href="https://civarchive.com/models/8821/extreme-realistic-tentacle-bulge">tentacle bulge</a> gives interesting shape to the bulge, and <a target="_blank" rel="ugc" href="https://mega.nz/file/MfB0kQCI#g8uxq7BEZk-YrSjCsqpwEl4auS3xiBYtGboXY-haBHo">autofellatio</a> (and maybe other futa ones) helps the AI draw absurd bulges by thinking it might be a dick at first.

hyper penetration stomach bulge

I've noticed v6 changing the artstyle on some (usually smaller) prompts so I lowered the learning rate, seems better overall.

Files with SHA256: 0d5eb826e6f713a388bd21bc3d7239be58a8f2e1598c9482d1a012ae55705fcc