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 about a week and a half on a 3060 12Gb. Cherry picked images for CivitAI. Generations are 0-shot with one adetailer face pass, because the model generally favors generating compositions with faces small enough that 512px generations struggle (classical paintings, methinks). No inpainting, negative prompts, no extra networks. No hiding behind the wonderful Rev Animated 1.5 finetune this time, not that I didn't give it a shot for added stability. This time I trained from base 1.5, and the params it came out with after training on the TFG dataset didn't benefit from merging like v0.6. Dunno what that means but... okay. Went from around 3k to 8k images, but quality floor and tagging accuracy are worsened. This model is NOT a reliable workhorse. However, if you want to see 1.5 do some things it probably hasn't before, TFG can probably do that now. I have tried to remove all artist and character tags, as personally, I'm more excited about models that generalize and understand style prompts, instead of just ripping a very specific style when prompted. If you want that, I guess try it with extra networks like LoRA.

Files with SHA256: beec1e9146b87619ec7a06ce5b95d2f257ef1534feebb3045e024558354503e8