一个Qwen edit稳定器，主要用于编辑图像生成后的图像与原图构图保持一致，
===========提供下载是你们自己本地用，不要kuku往其它平台上传，欢迎大家监督，不然没意思了，其它平台我自己不会传么?====================
使用说明”
Qwenedit稳定图像lora的使用说明及电商产品白底图生成背景的实测_哔哩哔哩_bilibili
kontext 版本说明：
kontext稳定器电商背景lora_哔哩哔哩_bilibili
关于产品的白底图，可以通过调整白底图布的构图留白程度，配合提示词完成摆放位置
降低权重可以使稳定作用降低从而不同程度的可以改变原图可用于一些需要创意改变的场景。
更多用法可以尝试一下，主要功能就是保持输入图的像素极的细节还原
最新demo，put it here 的封号佬帮研究出补充花纹的用法。~挺好
在线运行看效果及工作流下载：
基础用法
https://www.runninghub.ai/post/1967962698203934722/?inviteCode=rh-v1284
电商白底图自动（手动）生成场景图
https://www.runninghub.ai/post/1967977222789279745/?inviteCode=rh-v1284
修复图案
https://www.runninghub.ai/post/1967985030410678274/?inviteCode=rh-v1284
更多玩法发现请告诉我（最后有QQ，我发现也会更新）
训练目的：
qwen编辑后与原图不一致的问题
素材：
120对共240张对比图
用fill局部重绘制作
"rank": 128 因素材较多选了128 可能还是选多了，文件就比较大
V1训练步数：
28960步
A Qwen edit stabilizer, mainly used for editing the generated image to maintain the same composition as the original image, with secondary uses to be added later, in conjunction with the demonstration workflow. Let's send it out first
Training purpose:
The issue of inconsistency between the edited qwen and the original image
source material:
120 pairs with a total of 240 comparison images
Rank ": 128. Due to the large number of materials selected, 128 may still be too many, resulting in a larger file size
V1 training steps:
28960 steps
示例提示词：
editpicforgood,Only modify the requested content, and try to keep the rest of the image unchanged as much as possible,
背景变成沙滩，女人穿着鞋子
editpicforgood,Only modify the requested content, and try to keep the rest of the image unchanged as much as possible,
头发变成金黄色
kontext 版本的说明：
kontext稳定优化器及电商换背景lora | Kontext Stable Optimizer and E-commerce Background Change Lora
kontext稳定器电商背景lora_哔哩哔哩_bilibili
onlinetest with moblie
https://www.runninghub.ai/ai-detail/1970050962796802050/?inviteCode=rh-v1284
with web and download workflow:
https://www.runninghub.ai/workflow/1970051115515584513/?inviteCode=rh-v1284
在线手机测试
https://www.runninghub.cn/ai-detail/1970050962796802050/?inviteCode=rh-v1284
网页运行及工作流示例下载
https://www.runninghub.cn/workflow/1970051115515584513/?inviteCode=rh-v1284
kontext 训练到 33000步，好用一些，kontext本身画面是稳定的，这个lora主要是改善画面修改的质量，比如原先让它改变头发，它可能连眉毛也改了，头发的颜色也很奇怪，加了lora就舒服自然多了，可以试试看。
第二个使用场景就是和Qwenedit一样，电商产品白底图更换成场景图
原来1w多步效果不行，这两天加强到3W多步，好用了，可以和Qwenedit换背景有的一拼了
注意： kontext的触发词是editpicforgood 和qwenedit有区别
不过，不用触发词也是生效的
Kontext has been trained to 33000 steps, which is easier to use. Kontext itself has a stable image, and this Lora is mainly used to improve the quality of image modification. For example, if you originally asked it to change its hair, it may have even changed its eyebrows, and the hair color is also very strange. Adding Lora makes it much more comfortable and natural. You can give it a try.
The second usage scenario is similar to Qwendit, where the white background image of e-commerce products is replaced with a scene image
The original 10000 plus step effect was not effective, but it has been strengthened to 30000 plus step in the past few days. It is now easy to use and can compete with Qwendit for changing backgrounds
Note: The trigger words for kontext are editpicforgood and kwenedit, which are different
However, it still takes effect without triggering the word
我的qq 555649 欢迎反馈
Description
trigger words editpicforpartV1
but can run without it
FAQ
Comments (19)
文件好大
大的好，哈哈
Bigger does not mean "better" for LoRA models it usually means the LoRA is STACKED with zeros and unused weights. This is a VERY common misunderstanding.
@SencneS 你说的对，我设这个大小是因为我图片有1600张，800多对，不知道合不合适
@SencneS 欢迎给我更好的建议哈，感谢支持
@shendumao_4090_48g You're only training the model to keep the subject unchanged and not anything else, no new style, no new identity, new object, nothing. A giant lora does not make any sense for this, nor does a huge dataset and extreme steps. Using a high lora rank and a lot of steps both mean the lora is extremely likely to be overfitted to your training data and does not actually learn the concept you're trying to teach it.
@poyen Okay, this is also my learning process. Mistakes are the way I came here. I'll know next time
"最新demo，put it here 的封号佬帮研究出补充花纹的用法。~挺好"
The user who made "Put It Here" was banned?? It still seems to be here... what happened?
i still can see him ,Futurlunatic Creator Profile | Civitai
@shendumao_4090_48g Yes. I guess I got a bad translation. I thought it said he was banned. Sorry for the confusion.
just a chinese nick name
Партия гордится тобой. Партия выдать Buzz, и кошка жена
Работает! менять одежду можно легко
У меня есть lora, которая специально предназначена для белого негатива одежды, а затем становится моделью, чтобы носить его, но только в одном месте, потому что я учу потребности реальных продавцов электронной коммерции, если меня интересует позже я опубликую тест
@shendumao_4090_48g Хотелось бы попробовать вашу лору по смене одежды, эта лора работает но неучитывает тени, свет от фона
大佬，想要一个局部重绘的流，带遮罩那种，还有双这招迁移的，最好再来个参考图片风格转换的~
直接说功能目标，举个例子，衣服还是什么，说不定不用那么复杂
主要是局部重绘和万物迁移那种的，不区分衣服还是物品、人物，像fill那样加上遮罩，可以指定局部修改，多棒呀！参考图片风格转换就是，上传一张参考图，然后把原图转成这个风格的那种，嘿嘿！
@Just_hobby 风格转换按一张图其实不太容易，要看模型有没有类似的数据能力，之前的ipa其实效果也不会达到预期的效果，最理想还是要训练，想像一下一张图就可以代替一个花很久训练的风格lora，是不太现实的。毕竟 一张图里的色采，构图，等信息太有限了，局部修改用qwenedit的局部修改工作流试试喽，fill就不要用了，有点旧了，主要不太支持中国好多东西，比如你要换成包子，他都不会画包子，饺子