new version now available：
https://civarchive.com/models/104096/ffxiv-auraorff14-orff14
这个是关于FF14中敖龙族(晨曦之民)的LORA。
This LORA is made based on the "Au Ra(Raen)" race from the game Final Fantasy XIV.
肖像only | Portrait only
Recommend settings|推荐设置
Model: AOM3,OrionMix,Anything V4.5
VAE: Orangemix
LoRA Strength: 0.6~0.9(1 is also available)
size:640×640
Sampler: DPM++ 2M Karras
Sampling steps: 20
CFG: 7
Negative embedding: EasyNegative
模型：AOM3，OrionMix，Anything V4.5
VAE：Orangemix
LoRA强度：0.6~0.9（也可以选择1）
图片大小:640×640
采样器：DPM++ 2M Karras
采样步数：20
CFG：7
Negative embedding: EasyNegative
这个LORA有两个主提示词:au ra\(raen\),light-scales,还有可控的提示词:backwards horns,forwards horns,wings horns,downwards horns(控制4种不同的龙角生成)。
This LORA has two main trigger words: au ra\(raen\) and light-scales, and four controllable trigger words: backwards horns , forwards horns , wings horns , downwards horns (which control the generation of four different types of dragon horns).
backwards horns|forwards horns|wings horns|downwards horns
可以用:looking for viewer,from side,from behind 控制角度
You can use these prompts to control angle:looking at viewer,from side,from behind
>_<
Description
FAQ
Comments (4)
Nice, I've been waiting for a Raen model - especially with the wing horns. Too bad it's portrait-only unlike that Xaela model but anything is better than nothing. 🤭 Of course if full-body (with scales) is ever an option, that'd be the next step.
The training dataset produced this time is relatively small, with only 30 game screenshots. The overfitting of the art style is severe. To address this issue, I am currently increasing the size of the training dataset and improving its generalization ability. >_<
@hunix Nice. I'll say for keywords on the horn types, might help to give them unique names. For instance, instead of "wing horns" you could use "whorns" so it doesn't try adding wings as well. I was seeing wings starting to pop up on backs and I think so long as "wing" is spelled out, it'll try to add some.
@Eluna yea, I will try it. I previously used horns (type a/b/c/d) to distinguish, but the effect was not very good, possibly due to the small size of the dataset. I will try again this time.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.