DirectAlign Mitigates Reward Tampering LoRA

This is an early-version LoRA that adopts the DirectAlign* method; its performance is not guaranteed to be optimal.

This model is trained based on the ILLUSTRIOUS-v0.1 model.

The magnitude of the model weights has no correlation with image generation—you can use it right after loading (during my testing in Forge WebUI, I found that even negative weights work).

You are welcome to use it in combination with other base models and LoRAs.

Hope you enjoy using it!

GitHub link https://github.com/myAiLemon/FALumina

CivitAI link  https://civitai.com/user/AiLieLemon

LibLibAI link  https://www.liblib.art/userpage/7c113cc195604446a85f9b79eb973742/publish

DirectAlign*: Used to mitigate reward tampering in models.

---

It is prohibited to engage in behaviors such as merging this model for sale or directly monetizing and selling it.

这是一个使用了DirectAlign*方法的早期版本lora 效果不保证能做到最好

此模型基于ILLUSTRIOUS-v0.1模型训练

模型权重高低与生成图片无关,加载即使用(我在Forge WebUI测试 发现权重为负数也可以)

欢迎搭配其他基础模型&LoRA进行使用

祝用的开心

GitHub项目链接 https://github.com/myAiLemon/FALumina

CivitAI链接  https://civitai.com/user/AiLieLemon

LibLibAI链接  https://www.liblib.art/userpage/7c113cc195604446a85f9b79eb973742/publish

DirectAlign* : 用于减轻模型存在的奖励作弊

---

禁止对此模型进行融合贩卖,直接货币化贩卖等行为