LoRA trained on Blood Moon Elise from League of Legends
requested by: @Ritts
Blood Moon Elise: blood moon elise, 1girl, spider girl, short dress, corset, mask on head, short hair, white hair
models used
anime: YabaL_Mix - V5, MeinaPastel - V6 ( Pastel ), Sweet-mix - v2.0-flat,
2,5D: Animesh - Animesh-Pruned V2. ReV Animated - v1.2.2, PastelDiffusedMix - V2.1
realistic: Paragon V1.0
negative embeddings: bad_prompt, negative_hand, ng_deepnegative_v1_75t, bad-artist, EasyNegative
Recommended settings:
Make sure you are aware on the usage instructions of LORA
VAE: vae-ft-mse-840000-ema-pruned (for realistic models & 2,5D), kl-f8-anime2, blessed2.vae (for anime models)
Sampler: DPM++ SDE Karras or DPM++ 2M Karras(euler a for img2img upscaling) from time to time im also using DPM++ 2M alt Karras
Steps: 20 to 35
Clipskip: 2
Upscaler : GitHub - pkuliyi2015/multidiffusion-upscaler-for-automatic1111 4x-UltraSharp(for 2,5D & realistic), 4xAnimesharp(for anime)
CFG Scale : 5 to 9 (mostly 7)
LORA weight for txt2img: 0.5-1 (for the examples i mostly used 0.8 and 1)
Denoise strength upscaling: 0.3 to 0.55
If you want to autocomplete the tags & trigger im using, I recommend using Booru style tag autocompletion for AUTOMATIC1111's Stable Diffusion web UI
For full body shots i recommend using: ComfyUI Impact Pack
or the a1111 alternative GitHub - Bing-su/adetailer
Description
FAQ
Comments (5)
This one is so perfect with each of the models I use, amazing work. If you could, I'd suggest a lora for Quinn and Heartseeker Quinn since the only Quinn model is a checkpoint. I tried to scrounge up enough images to try and make a star guardian Quinn with my scuffed lora method, but there's just so few drawings of her sadly.
Sure, can do it. Shouldnt Star Guardian Quinn have more data then Heartseeker?
Why scuffed though? xD
@AhriMain I would've thought so too but it feels like a bit of a terribly neglected skins in terms of artwork, from what I found out least. Feel free to see if you can find a good pool of data yourself though because I'm sure a star guardian quinn lora would be absolutely amazing to see.
I used the kohya lora fine-tuner colab and my two attempted character loras of base seraphine went from v1 being a horrible disaster, to v2 being really good and nailing her outfit, but having way too much style bleed and broken hands despite bad hands negative embedding. Any seraphine model that has base is either a realistic style or a 3d cgi model type style so I tried my hand at making one myself.
Don't even get me started at my attempts at a style or concept lyco/lora lol
@dimpad I don't really know much about Colab, but I think you may have overtrained it a bit. I intentionally did the same thing to include more outfits, but it didn't turn out well, haha. There are two ways to approach this: either train with low steps and a high learning rate (0,0005) or high steps and a low learning rate (0,0001). I usually opt for the high learning rate route because it's faster, and it doesn't really make a difference when you only have one outfit. However, when you have different outfits that look similar, it changes everything. In this case, you need to delete similar tags(pruning) and use a lower learning rate.
My Step count is usually around 2000-5000, depending on the number of outfits or complexity
For Style LoRAs, you typically need a fuck ton of images. Tagging is not as crucial in this case. You can gather around 100-200 pictures and then use the wd 1.4 tagger to process them. Low/Mid Leaning Rate would be helpful here
@AhriMain I'll definitely have to give it another shot using your advice. Most of the time I'd just be going by whatever guide I could find said, and wouldn't know exactly what to do when something went wrong.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.