Yo. Its been a while i was having a very hard time getting results better than AIOMonstergirlsV3. I had trained merged trained merged again and so I took a break but now im back at it. This version has improved the quality of the image by a lot and also added two more monster girls that where requested.when training theses version i balanced the dataset making it so Lamia wasn't over-trained and Scylla under-trained. Also I refined the dataset by choosing quality over quantity and trimmed down on the captions. The style has change a bit again going for a more 2D look but with prompting and CFG you can get that 2.5D look again<ul><li>AIOMonstergirlsV4: The main checkpoint which is just a improvement from the base model with good overall concept comprehension</li><li>AIOV4Lora/lyCORIS: An extracted LORA from V4 about 90% accurate to the full checkpoint you will see some discrepancies it it about 340MB as its trying to put a full checkpoint and all the monster girls into a LORA</li></ul>one thing that has changed is I still recommend ADetailer but Dynamic Thresholding isnt really needed anymore and great results can be had with low steps and little tweakingAIOMonstergirlsV4 new monster girls:<ul><li>Frog girl</li><li>Mantis girl</li></ul>The LoRA/LoCON version is a LoRA not a checkpointFeel free to comment what other monster girls you want in for the next version. Most monster girls from Monster Girl Encyclopedia are somewhat trained but would need more training to get the quality i want. Have fun generating

AIOMonsterGirls

Complete change of style as I really like the saturation and shading of this version more

aiomonstergirls_v3.safetensors

AIOMonsterGirlsV3

AIOV4-LoRA/LyCORIS

AIOMonstergirlsV4

AIOV2-LoRA/LoCON

AIOMonsterGirlsV1

AIOMonsterGirlsV2

Moving to Civitai from hugging face you can find the rest of my models on my hugging face as i move them to Loras to Civitai as Loras didn't exist when I made thosehttps://huggingface.co/scricheAIOMonsterGirlsV3 is very good at Lamias and decent at everything else I over trained Lamias a lot and under trained the others one (especially Scylla). but still generates good monster girls i refined the dataset alot so i wont be able to generate all MGE girls nearly as well sorry but its for the qualityI know i said i would make LORAs for all the models but.... I suck. At making loras that is, so im just going to stick to what i know and make checkpoints and extract them :). Ill upload V3 Loras soonIve changed up the style of AIO for V3 as i really like the realistic shadow but still 2d look and i like my monster girls 2D :). This version is also way better at NSFW generation than the last two so go crazy. For the generated images im using a combinations of 2 extensions: Dynamic Threshholding and Detection Detailer the former is used for high detail without weird artifacts (usefull for the Lamis scales) and the ladder is great for monster girls as most of the detail go into the monster girl and DD automatically in-paints the face at high detail and fixes that almost good generationThe LoRA/LoCON version is a LoRA not a checkpointFeel free to comment what other monster girls you want in for the next version as most monster girls from Monster Girl Encyclopedia are somewhat trained but would need more training to get the quality i want. Have fun generating

Files with SHA256: 9ee2133245400d45dc6d92e682189f86550272f27bf7bc0b1e956ec0a6a17207