Hi folks!

I bring to you a LoRA model of my beloved Fizz from League of Legends. It is my first model, but I think the latest two versions turned out quite okay.

Looking forward to see more images from you people!

Versions:

v1.0

This was my first LoRA and I think it is a bit overtrained, so you get very unstable results. Generally would advice to go with v4.0.

Recommended Settings:

• Checkpoint: Any 2.5D does relatively well (tested with rundiffusionFX25D, Dreamshaper7 or ElegantEntropy_v12)

• Clip Skip: 1 or 2 (doesn't seem to make a big difference)

• CFG Scale: 3.5 (anything over 5 seems to be distorting the image a lot)

• Resolution: 512x512

• LoRA Weight: 0.7-0.8

• Hi-Res: Careful with steps and denoising. Works good on 4-5 steps and 0.4-0.45 denoise.

• No Restore faces

• Trigger word: Fizz (League of Legends)

Issues:

• Unstable body proportions

• Washed out colors

• Completely unpredictable tentacle hair

• Hands

• Feet

v2.0

Generally a lot of improvements in face and body proportions. Tentacle hair is still more of a miss than a hit, but sometimes you get lucky. Will probably try to work on this issue next. "full body" will render feet which actually often look quite decent.

Recommended Settings:

• Checkpoint: Any 2.5D does relatively well. Using Dreamshaper7 has been a blast with this version.

• Clip Skip: 2

• CFG Scale: 5.5-7

• Resolution: 512x512, 512x768

• LoRA Weight: 0.6-0.7

• Hi-Res: Don't go over 0.55 denoise, otherwise you are good to go.

• No Restore faces

• Trigger word: Fizz (League of Legends)

Issues:

• Unstable tentacle hair

• Unstable hands

v3.0

Improved flexibility of the model and much more stable tentacle hair and hands. Improved NSFW generation.

Recommended Settings:

• Trigger Word: fizz_lol

  • Additionally use tentacle hair to give the hair generation a slight push into the right direction. I often found a weight of 0.7 is enough.

• Suggested Weight: 0.6 - 0.8

• Please refer to the gallery for all other settings, as they can wildly differ depending on prompt, checkpoints and other loras.

NSFW:

NSFW can be triggered by using cetacean penis and/or nsfw in your prompt. My experience with nsfw has been, that it introduces more weight on the base model, which often led to generations that were not resembling Fizz.

v4.0

PLEASE NOTE:

• Reinforcment tag for tentacle hair has been changed to tentacle_hair

• Reinforcment tag for cetacean penis has been changed to cetacean_penis

Major improvements to generation stability thanks to the Scrimblo Sauce Furrymix basemodel, but at the cost of having a harder time to generate hyper realistic images. You'll also appreciate the much better NSFW generation.

Recommended Settings:

• Checkpoint: Most furry checkpoints should work fine

• Trigger Word: fizz_lol

  • Additional reinforcement tags:

    • webbed feet

    • tentacle_hair

    • gills

    • short tail

• Suggested Weight: 0.6 - 0.8

• Please refer to the gallery for all other settings, as they can wildly differ depending on prompt, checkpoints and other loras.

v5.0

Again major improvements, this time thanks to the PurrfectMerge checkpoint, provided by PurrfectlyAI. All provided showcase images are raw out of the box generations with no hires-fix, img2img, inpainting or other improvements. Version 5 generally tends to generate 2.5D/3D images, but using the right checkpoints and prompts can still give you nice anime/drawing generations.

This versions changes include among other things:

• General stability improvements, less artefacts or weird appendages

• General versatility improvements. Can be used with more different checkpoints. As long as they are for furries you should be golden.

• Generations have much better likeness to Fizz, e.g. body proportions, colors, body markings, skin texture, etc.

• Improvements to hand and feet generation

Unfortunately, underwater images struggle with the tentacle hair for some reason and I also failed to introduce his trident. Haven't found the right approach for that yet and at this point I am inclined to let it go.

Please refer to the gallery for settings, as they can wildly differ depending on prompt, checkpoints and other loras.