[BETA] Riot Diffusion | League of Legends style (Read info!)

Trigger words: LOLSPLASHART <- Always use it! Model made by user: xxxxxxx. Trained with 4090 on all League Splash Arts in NAI model. Try to use it with face fix. Link: https://civitai.com/models/81758/lolsplashart-league-of-legends-splash-art-base

How to use it: Use the activation word, the name of the character and optionally the name of the skin. Don't forget to include the activation word or the model won't work correctly. Also don't forget to name the skin. Example: lolsplashart, jinx star guardian, 1girl, long hair, breasts, looking at viewer, smile, open mouth, bangs, hair ornament, red eyes

Recommended use by the original author: Resolution 1024x. DPM++ SDE Karras with 20+ steps for the best quality (If you use Euler a, try 27+ steps). Face fix + HighRes fix with 0,55 to 0,6 denoising strenght (mostly using 0,58 - I'm not sure about HiRes steps but it looks like +30 is good). CFG 5. The author says it doesn't generate good anatomy all the time.

Uploaded by darkskyx. Note by me: The official resolution for all the LoL Splash art would be 1024x614, but the author is using 1024x576. Additionally, you can try it with my LoL LoRA (from 0,3 to 0,4 weight with or without the LoRA activation word): https://pixai.art/model/1702471936327154878 If you use the activation word of my LoRA, you can also not use the activation word of the model, so my LoRA will be more powerful. You can also try other DPM++ samplers. In my opinion, this model is not 100% suited for PixAI and it creates bad hands. You'll use too much credits.

Recommended by me for a cheap generation: Use Euler a with 20-27 steps and my splash art LoRA. Then enhance the image with another similar model to give it better hands and body. I also recommend to use a LoRA that increases quality/details so you don't have to use x1024 resolution. (Optional: If you are free user, erase the LoRA for quality if you have used it and while enhancing, add a LoRA for hands).

Since this model has been trained with few images, it seems that it is better to first generate this style and then maybe try to fix it with other model with better anatomy... or with the own model with some luck.