This lora was trained on about 60 images from the artist Nagi Itsuki. The images used were mostly backgrounds with a few characters in between. The lora captures the style and shading but fails miserably to replicate the details of the source images backgrounds. You will notice this if you attempt to prompt a large scenery. I feel this is a limitation of SDXL more than anything.I recommend you stick to close-up and simple concepts. The character style replication works well enough but it becomes influenced by the model once you begin to prompt concepts outside of the artist's territory. The actual character art of the source is much more simple than what you'll get from this lora due to the model. If you want something more simple similar to the artist's character style, I recommend using less score tags.This lora was trained on AutismMix and I haven't tested on anything else.To get the results you see in the images above, these are the settings I used.Prompt(at the start): score_9,score_8_up,score_7_up,score_6_up,source_anime, lora:Nagi_Itsuki_XL:1 [your prompt] (You can use "scenery" as a keyword but it's quite strong)Negatives: mosaic censoring,bar censor,censored,source_cartoon,source_pony,source_furry,simple background,blurry,blurry background,fading,Sample: DPM++ 2S aSteps: 26 (Whatever gives you the best results)CFG Scale: 7Extension: [sd-webui-color-enhance] set to 0.5 (this increases the saturation a bit)Upscaling: Follow this methodI exclusively use AutismMix and I recommend you do the same. All of the examples were made with that model.
