Ive learned a lot making my Naoto lora and applied everything I've learned to remake this lora and I'm very happy with the results. Trained on NAI with 118 images of dxd gacha + some anime screencaps. It's V4 as I've retrained and carefully finetuned it several times until it was perfect. Spent another week on this one, all reviews are very appreciated. Looks good on every model I've tried except animelike, that model just cannibalizes the style and character design. The outfit is pretty consistent but not perfect as the triangle extensions don't always appear. All sample images generated with high res fix.
Tags for main outfit I highly recommend: konekouniform, school uniform, white shirt, skirt, red skirt, black ribbon, black corset, kneehighs, black socks,
Tags for Koneko: Koneko Toujou (DXD), white hair, short hair,
描述:
-
Trained on NAI instead of Stable Diffusion 1.5
-
This lora is no longer a generic cat girl and is now consistently recognizable as Koneko and consistent with her body type
-
Main outfit actually works well and is consistent
-
DXD style is now more accurately portrayed
-
Redid all tags in the dataset for more consistency and reduced bleeding which fixed many of the V1 issues with frills/bows/bare shoulders/etc being baked into the model
-
Base activators for main outfit changed to - konekouniform, school uniform, white shirt, skirt,
-
Removed the cat hair ornament from the dataset so it no longer appears
-
Now works perfectly well at 1 weight instead of 0.7 like before
Parts that are worse:
-
Cat tail may be worse and simply come off screen even when not prompted
-
The style is stronger and may override the checkpoint's style a bit depending on the flexibility of the checkpoint
-
Cat ears and tail may sometimes generate black for some reason
-
Because it's trained on NAI which is trained off Koneko already, it sometimes places card frames/text on some generations even though there isn't any in the dataset
-
Gacha card backgrounds appear more often, but tagging a specific background usually will get rid of them
Trained on NAI instead of Stable Diffusion 1.5
This lora is no longer a generic cat girl and is now consistently recognizable as Koneko and consistent with her body type
Main outfit actually works well and is consistent
DXD style is now more accurately portrayed
Redid all tags in the dataset for more consistency and reduced bleeding which fixed many of the V1 issues with frills/bows/bare shoulders/etc being baked into the model
Base activators for main outfit changed to - konekouniform, school uniform, white shirt, skirt,
Removed the cat hair ornament from the dataset so it no longer appears
Now works perfectly well at 1 weight instead of 0.7 like before
Parts that are worse:
Cat tail may be worse and simply come off screen even when not prompted
The style is stronger and may override the checkpoint's style a bit depending on the flexibility of the checkpoint
Cat ears and tail may sometimes generate black for some reason
Because it's trained on NAI which is trained off Koneko already, it sometimes places card frames/text on some generations even though there isn't any in the dataset
Gacha card backgrounds appear more often, but tagging a specific background usually will get rid of them
训练词语: koneko toujou (dxd),white hair,short hair,konekouniform,school uniform,white shirt,skirt
名称: Divine_KonekoToujouV4-000010.safetensors
大小 (KB): 18545
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
