Arcanium Aurora - [NoobXL] - trained on YiffyMixV63

詳細

モデル説明

This is a LORA aims to generate character calls Aurora in the roguelike card game Arcanium: Rise of Akhan by Supercombo, Inc.
Aurora herself is a pretty generic anthropomorphic fox character, the only things that make her stand out are her outfit and maybe the early-2000s art style.
On a side note, the game is a bit rough around the edges with its balancing; there's a broken debuff combo that completely trivializes the content and overshadows all other tactics in the game.
However, this game is a steal at 75% off, you'll get a solid 20+ hours of gameplay out of it.

All showcase images are cherry picked and generated using YiffyMixV64 checkpoint, enhanced by Hires. fix with 125% upscale and 0.75 denoising strength.
This LORA has a glaring issue with heavy artifacting on fine details (especially the eyes in standard 1 million pixels resolutions, eg 832x1216, 1024x1024 and etc). While I have yet to find a solution, utilizing Hires. fix can mitigate this issue to a certain extent and yields acceptable results.

Below is the ingame Aurora character image


Trigger word: arcaniumchar_aurora

Recommended tag: red long sleeves open shoulder kimono, digitigrade leg armor, stockings, gold shoulder guard, big tassel belt, blue long waist sashes, blue loin cloth, vambrace, necklace, bracelet

Recommended LORA Strength: around 0.75

Prompting Style: e621 or danbooru tags, requires the recommended tags to manifest the Aurora's outfit

CFG strength: 4~5, higher values tend to fry the fine details of the generated image

Post-generation Enhancement: ADetailer and HiRes.fix can help greatly in reducing the artifacts around the eyes


LORA description
Trained on YiffyMix V63 checkpoint with the Rank/Alpha ratio of 16 to 8.

LORA known weakness:
1. Serious eyes artifacting: Because the source images are low quality, the LORA struggles to generate clean eyes at the standard 1 million pixels resolution. Using post-generation enhancement like Hi-Res Fix can significantly help to mitigate this issue.

2. Tendency to generate blue flame unprompted: you can minimize this unwanted effect to some extent by adding 'blue flame' to the negative prompt.

3. Finer details not being generate accurately: the twin tassel belt might turns out looking weird or end up a total mess.

4. Unable to generate style accurate character: This LORA captures the character's likeness, not their original art style. Don't expect it to generate style-accurate images.

このモデルで生成された画像