Modders locked model to "intended to generate mature content". Big L for everyone with blue buzz, sadly. :( Base:Another LoRA for (surprisingly) underrepresented concept. Main goal is to produce images, where relief of bulging intestines (booru tag "<a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=intestines_filled">intestines filled</a>") is visible on abdomenal region of a character (most often - from tentacle penetration, but also can be from very long dildo, inflation with fluid etc.). Dataset will be mostly synthetic, as it is surprisingly hard to fish out good images with this concept. v3.0: 49 images in dataset. Trained locally with citron-trainer (wrapper around kohya-ss). caption was made with 'SmilingWolf/wd-eva02-large-tagger-v3' and 'huihui-ai/Qwen3-VL-4B-Instruct-abliterated', where tags were merged from both. (due to mistake in code <a rel="ugc" href="https://colab.research.google.com/drive/1Xf7CKhamMuuySmu14FtyBQXwQO9OCt_z?usp=drive_link">here</a> blocks of tags and NL was reversed and I saw it only after the fact. Don't know, how this would affect results, but probably negatively). More new control words, but "colon filled" was lost in transit. Shape not appears at random. Biased towards exposed belly and general nudity. Due to lack of diversity in dataset affects style (drift towards slight blurr and sepia). Weird, bad looking transitions frrom bumps to skin around are still present.v2.0:Next itteration. Used 41 of synthetic images, standard civit trainer settings. Tried to add new concept of "colon filled", but it seems that this whould require clip training also. Currently both concepts are present but still unstable.Overall quality seems to be worse, but this is most likely connected to bad tagging/captioning (it is not smart from me to do such responsible task at 4 o'clock in the morning).v1.0:First round of training. Dataset of 15 images from internet and 1 photoshopped picture. Default CivitAI trainer settings (850 blue buzz spent). Dataset has both captions and tags in different amount, done by hand (tried to tune captions to produce needed image minus bulge itself, if used as prompt for anima). Dataset available, but may be removed if someone get feisty about used images. Results are unstable, but already have some shapes in right places. Lora also supposed to make distinct bulge, inflated stomach and filled intestine in different combinations. Most images in dataset has tentacles, so lora also generates them. Results from this lora will be used to train future versions.

Intestines filled (inflated/stuffed/bulging)

Files with SHA256: e2779ce8eb4b2a0b5b3ca1c1fa87cd2bb2cd51bd3e93119364824913d3b0e326