Hellhound
V5, fixed the overfitting, Single Triggerword "Hellhound", Her skin fluctuates from Black to Dark Grey, every once in a while her hair might turn grey, if it does I just turn the strength down to .9 or generate again depending on the complexity of the prompt, backgrounds are much better now and the odd glowing effect is gone. Trained with a .5 network dropout and that seemed to make her much more generalized and able to adapt to a wide variety of prompts while still keeping her basic anatomy. Now only using 1 tab On AI Detailer unless I want something specific to have it pay attention to other than the face/eyes. Capable of all manner of clothing, figured out that adding score_9 and source_anime to the tags forces the training to learn better through some unknown wizardry and the results speak for themselves. Occasionally if using with other lora's it might need additional anatomy tags if her anatomy fails to generate, doesn't happen all the time but it happens and I can't predict compatablity with every other possible lora, maybe If I can get enough input images and train using this new method I can increase it, will continue trying to improve her as she is my favorite character.Version 4, Consolidated into one Trigger word, "Hellhound". 100% increase in ease of use, Her black/grey black skin color now generates without prompting for it, if you want different skin colors I recommend turning down the strength to .7 and prompting for it Seems very reliable on autism mix which this was trained on, Base Pony not so much as the hair color can break quite often. Eyes are a bit better but will still require inpainting on .55-.65 strength to look their best, mouth is the same, better with inpaint. The flaming eyes polluting the model have been eliminated at least during my testing, if you want them I suggest inpainting them otherwise they would look janky through standard prompting. Though they are no longer required in the negatives to avoid them.Tested with a variety of poses and clothes and didn't see any major issues. Strongly suggested to add "source_furry or source_anthro" to the negatives to avoid any random anthros, it's not very likely but is possible to generate. I probably over trained this a bit but I like how it looks. Once I learn more regarding the optimum settings I will continue trying to improve her. Highly Recommend AI Detailer as it removes the need for inpainting as a secondary part of workflow.Version 3 is here, trained using existing knowledge, activation tags have been changed"Hellhound, monster girl encyclopedia, black skin, black sclera, red eyes,"90% of the time this will generate her accurately with animal hands and feet, feel free to add those tags and "wolf tail" if it fails to show up or you require more precision.is capable of some 3d renders, though I am not a fan of it so I didn't test much of that aspect.Eyes will require inpainting since they are kind of rough, eye flames are toggleable to an extent, using negative prompt, sometimes wierdness will appear even if you leave them off since this uses the base models knowledge. So I prefer to leave them off and in the negative prompt. All in all I am very happy with how Version 3 turned out, learned a lot in the process through lots of trial and error. Many thanks to the creator of MGE KC and many fan artists.Feel free to post some stuff you make I want to see it, thanks and enjoy.[Version 2] outdated and inferior, Increased Images in Training Data Set, Retrained On Anime PonyXL to Eliminate Non Anime influence on the model, seems to have fixed the random anthro/furrys showing up in generation since Anime is the focus on this model, Ms Hellwan is much more consistent now. Recommend CFG 6-7 and any of the AutismMix models.
