Just put it into Positive prompt(V3.5) and it will work well！
use ESD for combine both positive and negative prompt.
positive：(masterpiece,best quality,ultra_detailed,highres,absurdres:1.2)
negative：(worst quality, low quality, ugly:1.4), poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, cross-eye, blurry, (bad art, bad anatomy:1.4), blurred, text, watermark,
Recommend Weight:-1~2
t2i：-1～2
i2i：0～2
Sample:
positive: 1girl,black hair,hair_flower,petals,<lora:neg4all_bdsqlsz_V3.5:1>
Description
differentce use with V1
just put it in positive prompt
FAQ
Comments (32)
Is it a LoRA or a LyCORIS?
C3lora ,it can use directly
Your preview marks this as LoRa but it is labbed as a LyCORIS, mistake?
no,it is C3lora and it can use directly...
it means locon in lycoris
@bdsqlsz It's not clear at all. What should I specify in the prompt - <lora> or <lyco>? What folder should I put it in - for loras or lycoris? Can you explain clearly?
@bdsqlsz this response clarifies nothing for me
@magenta both OK and use it any way you want....
@Merkuriij now locon can be used in sdwebui directly, it is lyco.
谢谢你，青龙
what this is for? dont c any difference in sample image between weights.
it combine these concept：
positive：masterpiece,best quality,ultra_detailed,highres,absurdres
negative：worst quality, low quality, ugly, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, blurry, bad art, bad anatomy, blurred, text, watermark,
我测试了4种情况，放在正面tag里权重为1和-1，放在负面tag里权重为1和-1，画面没有实质变化，感觉结果和之前那个修手的lycoris一样，这类微调方式可能不适合去修正画面，感觉真正导致画面问题的还是模型的参数量，训练集质量和数据标注的准确性。
新技术并不使用任何图片数据集进行调整，而是直接使用大模型的概念进行训练。
使用正面tag 3～4会有很明显的效果。
1～2也会有效果，但是不会非常影响画面本身的内容。
@bdsqlsz 先感谢青龙，你的所有Lora的教程我都有看，我能学会Lora也是感谢青龙的无私奉献，谢谢。
确实直接调用已有的概念进行强化会有更好的效果，我会尝试更高的权重，但Lora还是只是在注意力层做手脚，不知道mid是用了什么样的方法去改良画面的问题，之前有考虑过可能是利用增量的数据集对模型进行调整，但是我尝试了用db结合增量数据对大模型调整也没有办法解决手的问题，而且现在看来无论是Lora还是db,或者emb都没有办法达到midjourney那种对画面的改良程度，我觉得他们可能是对模型结构和训练方法做了改变。
@monster_maker 这里用的是25层的C3lora，包括transform也做了训练，实际上就是lyco的locon。
至少目前这个lora能让手指保持在5个...
试了一下，如果带着原本的各种质量词、反面提示词、文本反转模型的话，低权重修一下就足够了。但这些全去了最好还是高权重。谢谢龙佬！
看了B站动态，就马上杀过来了，先赞后用美德也。
试用了一下，权重从-1~4，总体感受还是比较挑底模的，有些很好，有些则完全不行，大佬是否考虑拉一个底模的推荐名单，这样更方便小白使用。另外有两个问题，一个是当微调参数时，每调整0.1个权重人物姿态都会有较大变化（已锁定seed）不知是否是底模问题。另一个问题是当叠加其他lora的时候，会出现明显的过拟合，不知是否有好的参数搭配的建议。
how can you put negative embeds into a positive? How did you make this?
sorry for replying late. it is ESD.
@bdsqlsz What's ESD? I tried googling it but found nothing
@bdsqlsz wow, so... I wonder if it would be possible to make a Lora that erases vaginas....
@Janet of cause it can....
@bdsqlsz I've been trying to train it to do this, but no success.. it would be so useful for me
@Janet u can use LECO on github
@Janet just vagina? i can make a SFW lora for this
@bdsqlsz yes, I would love one that could remove vagina! Removing long hair/breasts would also be nice, but that's asking too much. The main difficulty for me is vagina.
@Janet ok,ill do it when this weekend
@bdsqlsz thank you!!!