V2The images used for training were cropped and re-scaled to improve training quality (this helps to reduce the unexpected behavior of generating multiple characters in a single image even when the 1girl/boy tags are used). The style trigger word (duel art) no longer requires a weight greater than 1.I trained on WAI-NSFW-illustrious-SDXL v12.0 as base model; the showcased images were generated using that model.The images were generated in Krita using Hires upscale: 1.5V1LORA trained on the non-sketch images from the "Duel Art" artbook. Images were double page, so this LORA tends to create images with more than one character per image (this is my very first LORA, I'm planning on doing V2 with more a more appropiately processed dataset).The trigger word is "duel art". It seems to need a high weight to truly use the stlye (I used 1.2 as you can see in the sample images). The LORA was trained on vanilla Illustrious as provided by Civitai, the showcase images were generated using WAI-NSFW-illustrious-SDXL.The images were generated in Krita using Hires upscale: 1.5

kazuki takahashi duel art

The images used for training were cropped and re-scaled to improve training quality (this helps to reduce the unexpected behavior of generating multiple characters in a single image even when the 1girl/boy tags are used). The style trigger word no longer requires a weight greater than 1.

1769243_training_data.zip

duel_art_v2.safetensors

V2The images used for training were cropped and re-scaled to improve training quality (this helps to reduce the unexpected behavior of generating multiple characters in a single image even when the 1girl/boy tags are used). The style trigger word (duel art) no longer requires a weight greater than 1.I trained on WAI-NSFW-illustrious-SDXL v12.0 as base model; the showcased images were generated using that model.The images were generated in Krita using Hires upscale: 1.5V1LORA trained on the non-sketch images from the "Duel Art" artbook. Images were double page, so this LORA tends to create images with more than one character per image (this is my very first LORA, I'm planning on doing V2 with more a more appropiately processed dataset).The trigger word is "duel art". It seems to need a high weight to truly use the stlye (I used 1.2 as you can see in the sample images). The LORA was trained on vanilla Illustrious as provided by Civitai, the showcase images were generated using WAI-NSFW-illustrious-SDXL.The images were generated in Krita using Hires upscale: 1.5

Files with SHA256: b27a83715b66d5dd0aedaacb116fdcc79a24874112e72083a6ca576d70e6e542