TLDR: Better proportions in some scenarios(?), older median age, interesting outdoor landscapes, painterly vibes. NOT a reliable workhorse yet, still has lots of weird and bad. Be aware that the model will produce NSFW content, and may not be able to reliably adhere to prompts/give you exactly what you want due to dataset and tag noise. This model was tagged using vision enabled large language models trained on Danbooru's tagset. Dataset is a fairly large scrape. YOU HAVE BEEN WARNED, THIS MODEL LIKELY JUST PRINTS NIGHTMARE FUEL. The Female Gays is a dataset, and maybe in the future it will be a series of models. The name is a joke about lesbians liking older women, and wishing the biases inherent in anime (booru tag based) models had a more 'feminine' sensibility. SD1.5's initial anime finetunes had poor performance generating adult women, and LoRAs that might've remedied these issues at the time were overfitted, hyper fetishized models incapable of generalizing. I wanted female characters with more realistic body types and less bias to young folks when rendering 2d art. It wouldn't be useful to me if it only returned overly fetishized depictions of adults, meaning I couldn't just pull from a mature_female tag for fear of training a model to produce pea heads and hyperbreasts (and worse or better depending on your angle). Last, I wanted to maintain model flexibility so I could actually use this for my own SFW applications, like storyboard drafting utilizing LoRAs. I started wondering if it was possible to finetune a checkpoint on a consumer rig and fell down the rabbit hole. Roughly 90k images passed through in house image classifiers, narrowing the set down to around 7k using custom ML image classification tools and open source python libraries in tandem. CLIP, YOLO, Finetunes like deepghs/anime_censor_detection, Multiple OCR systems like easyOCR, resnets, yada yada. Around 40% of the removed images are candidates to be included in future iterations of this project, sorted and worth training on aside from a minor feature, but I have currently put WAY too much time into this! So there are like 25k sorted identified strong candidates that can be added with some time, allowing eventually for a better quality floor. Tagging also needs to be cleaned up for what we have already. Most of what I am doing is cleaning out the YUCK from the imageboard sets, and there is much more to do. Despite all that, I have been genuinely surprised a few times by this model, as it has produced things I hadn't seen 1.5 do before (my narrow experience). That being said, I hope this is at least novel to someone, and three cheers to artists who create works with diverse women in mundane situations.

The Female Gays - SD1.5

Trained for 4 days on a 3060ti. Over a week of straight training (SO hot in our tiny spot, s/o to the s/o) on the combined versions before this, so many hours of sorting and tagging. Need to be done this in some capacity, so enjoy this early release! num train images * repeats / 学習画像の数×繰り返し回数: 19645 num batches per epoch / 1epochのバッチ数: 9833 num epochs / epoch数: 30 total train batch size (with parallel &amp; distributed &amp; accumulation) / 総バッチサイズ（並列学習、勾配合計含む）: 2 gradient ccumulation steps / 勾配を合計するステップ数 = 1 total optimization steps / 学習ステップ数: 294675 Noticed problems: Hands suffering In extremely wide shots or where subjects aren't the focus, faces are distorted Pink skin Seems to be framed images hiding in the dataset despite over a thousand manual crops Too much style transfer from classical paintings Subjects often generated with antiquated clothing or hairstyles, needs better tagging to favour contemporary Dataset might include some weirdly specific scenarios in its dataset that force near regurgitation of training material Dataset noticeably not varied enough: needs more clothing, styles, poses, etc This is only speculation from playing with it, but it seems it really had to rip apart some things it had learned about anatomy and was only just starting to re-converge on things like faces, fingers, 2 arms, 2 breasts, etc. My x/y spreads seemed like errors are STILL decreasing with training, as well as the loss still lowering (ended at .083) After testing was pretty obvious, epoch 30 was merged back with Mama REV V2 Rebirth to try and lessen the effect of some of the problem points while still being able to observe the effect that the training had. The result seems to be somewhat stable! Maybe not! Either way future versions may either be trained from the SD1.5 base or will just rely a lot less on rev's vibe. Thanks a ton for the folks who have touched that project.

Files with SHA256: 4ea0777a9b115341356d5cb6f9f37b7ae7e80851e6821dee105d8d589193e5fb