The new version is for Klein9B. thanks to Klein's ultra-fast response speed, image editing has entered a new era. The training of this LORA was full of twists and turns, but I was lucky to get this version, which has completely surpassed my previous Qwen version. I call the new version "my nano banana moment". Friends who have experienced its performance in advance should understand what I'm talking about ( for those who are not sure, you can click here to learn more first ). I won't say more. I wish you all a great time!
Experience the A2R Sponsored Version here ( online workflow ) 👈
want contact me👇
X : @AIAVJLeo | [email protected]
| 🙈🙉🙊OK, that's all
If you like my work, please like, follow, and share. This will help more people! Thank you.
English is not my native language. The above content was translated by artificial intelligence. Please forgive me if there are any inaccuracies.
Finally, have fun! AI never sleeps! Peace!
Description
FAQ
Comments (71)
I have tried in several images, but it does not work well with lighting loras, 4 steps or 8 steps. I 'm not sure if is a problem with the CFG = 1.0, or the number of steps.
can I see your original image?
having a premium version of a lora is shitty. i hope no one buys that dumb stuff
Why do you say that? How did it offend you?
Cause its true, hope someone steals back all your money if only one image of your trainig data is not your personal property.
I personally just fixed the low quant issue in 2509, FOR FREE AND THE WHOLE COMMUNITY.
all following words are censored
He had to train locally on his modified 4090 a model for more than 30hours which actually works wonders compared to the low quality garbage released daily on this site. Then after releasing it for free without nobody forcing him to do so, he decided to train it even more and try to monetise over it, which all considered it's may not be a stupid idea. We get a great quality model, he (maybe) earns some bucks, so it's a win-win. Hopefully he will keep releasing stuff for free. Maybe he will also lower the price from astonishing 35$ to 10$ so people will buy it.
@Couldvebeenworse Thank you. I originally didn't plan to pay attention to him, but after seeing your message, I have to say that being understood is really a happy thing. Thank you for your support.
Maybe all the cool guys making it even possible to use/train models like 2509 keep their future stuff behind a paywall.
Maybe after that some of you understand what community means.
30hrs of training makes me laugh.
It took the last 3 month to check, write code, doing the same images over and over, to find issues in repos like stable-diffusion.cpp ComfyUI, City96, reading thousands of lines of code.
All of this guys spend their compute/brain power for free, since years.
And yea, in my mind, maybe i looked too much star trek, we are a community, happy to help each other, happy to learn from each other, happy to share to each other.
And again i laugh over 30hrs of training, because you use it for your own, you decided to go further, cause you wanted to do it better. Thats our inner Monk. And thats good.
if someone of you cant even afford a 4090 and needs to use a gguf of qwen-image-edit-2509.
Think about you had to pay for gguf support, maybe someone like me, just testing, having fun but better investing in my kids than in a capable computing rig. Would have never tried qwen-image-edit-2509.
then something like this probably never happened:
https://huggingface.co/QuantStack/Qwen-Image-Edit-2509-GGUF/discussions/6
if 10 ppl using my models i get maybe 3 of 10 constructive critics.
if 10k ppl using my models there is much more potential to get knowledge.
thats from what we all take and thats why we should give.
dont get me wrong, im personally not offended, i understand that you are a little bit proud and want to get a profit of your hard work but on the other hand it makes me kinda sad.
the other thing is if you share, normally nobody really cares and closes eyes. if you sell something youre official and if someone gets offended you could get charged eg. if there was only a single pokemon image in your training data.
a private person gets away with signing, hey i take it offline and never do it again.
a official seller gets a nintendo charge of 50k or similar.
i guess i should have used other words im my fist post to explain.
I want to make my own qwen lora, can you share details on what program you used and how you prepared the images?
Are the images two images in one, with the left the original, and the right the new image? Does the caption cover both images?
If you can help me with an easy to understand example that would be really helpful.
Thank you.
Have you done model training before?
@vjleoliu Yes, but for regular Flux not an image editing model.
@vigilence oh, It might be hard to explain clearly here. Why not check YouTube? The videos there will explain it in more detail.
@vjleoliu Did you have a video you would recommend?
@vigilence Sorry, I have always been focusing on videos in the Chinese-speaking region, which may not be of much help to you.
I am also training with qwen lora, but I have noticed that it is difficult for qwen to achieve a texture comparable to that of the training set merely through lora. Therefore, for the edit model, I mainly enhance the association with the reference image to achieve a certain conceptual reference
This is a very good way to use it. Could you elaborate a bit more?
Also, regarding training, my suggestion is that if you want Qwen-Edit to approach the texture of the training set, then let it overfit as much as possible. Take my LoRA as an example: the Base version used over 10,000 steps to make the texture more realistic, while the Plus version used even more steps, and the result exceeded my expectations.
It's interesting how it works very well if the original is very flat, even sketch-like, but an already semi-real'ish drawing barely changes at all. Hoping for a more general lora that can do both in the future.
Yes, the same situation will occur if the original image is a 3D image. It is speculated that this is related to Qwen's judgment mechanism. The closer the original image is to a real image, the more likely it is to cause Qwen to make a wrong judgment.
This is the current transitional processing solution. I have tested it, and the effect is not bad. I hope it can help you.
Is there a specific trigger sentence or words to use like in your examples "change the image into realistic photo"? Which was the base model prompt used as comparison? Are there specific resolutions/aspect ratios, Samplers/Schedulers your lora will work as it should?
Sorry, I seem to have forgotten to publish its output parameters. Maybe I should post some images with metadata on Patreon later, so that everyone can understand more clearly. Here is the answer to your question.
The trigger words for both the Base version and the Plus version are the same: changed the image into realistic photo.
The ratios I commonly use in tests are 2:3 and 3:4, but in fact, most common ratios are fine as long as they meet Qwen's image generation standards.
In ksample, I use euler/simple, but actually, using other schedulers works better. I use simple just because it is faster.
I hope my answer can be helpful to you.
with this lora it is alwyas a black image. the samsung lora works so it is definitely not my problem.
2509？
ok, so try disabling sageattention and see if you get something. That's what I tried and it is working now for me. I was getting just black boxes.
@eca132 Same, it's sage-attention at the command-line that causes this. No idea why.
@brnlittokhoes3110 I have no idea either, I saw this as a fix from another conversation about the problem... tried it and it solved it.
兄弟你完全是个天才！太好用了，我使用的是aio v5.2 效果很棒！
我擦，居然碰到同胞啦，你怎么知道我说中文
觉得好用就多点赞多返图，返图区冷清的很呢
@vjleoliu 不是哥们我还歪打正着了？其实只是因为我太激动所以懒得用ai翻译成英文。不过兄弟你这个模型有一点真的把我难住了，就是关于nsfw的问题。我的原图是仅仅漏乳晕没漏乳头的，但出的图一直存在乳首，而且qwen的乳头生成的真的特别难看，我提示词还禁止不了。
挑选了几张能看的。再次感谢你这个优秀的lora模型。
@RUmurs 哈哈，没事，多返图多点赞，推推这个模型，可以让更多人用起来。关于你遇到的问题，我很少尝试nsfw，所以也不了你太多，要不你试试局部重绘呢
@vjleoliu 哥们是我的问题吗，在我是设置了图片用了这个模型，想要返图但在此画廊下没看到我的作品 T_T
@RUmurs 可能是C站的问题，所以我通常都在lora页上发图
I tried several times but the result still be failed. 2509 edit-fp8, AR-2509-base + lightning 8X,prompt:changed the image into realistic photo.euler + simple
What does the original image you used look like? Can I take a look?
@vjleoliu I know the root cause from Qwen-2509 edit-fp8 now. When I change model to Rapid Aio, thing is normal. I am confused about Qwen-2509 edit-fp8 because I download it from official channel. I have same issue for "Next scene" lora under model of Qwen-2509 edit-fp8, I guess it has same root cause. Thanks for your kind help!
@yli6hz can you please elaborate how you are using the Rapid AIO checkpoint ?
i just want to drop a comment to say henlo, thank you for this awesome lora ! :D
lol，thx，bro，If you think this model is good, please feel free to upload your test results and give it a like. This will help more people discover it.
VERY effective . Prompting with details yields excellent results. Lazy 'trigger-only' prompts get mediocre results compared to being specific with what you want to see. The base model simply can't do what this Lora can. I've added samples, but the site isn't showing them all sadly. Thanks for sharing this fantastic resource!
thx.bro
@vjleoliu No Prob! One more note: This model/Lora combo is REALLY stubborn with faces. It decides what the character should look like, and barely changes with many prompting attempts (describing face features, ethnicity, etc). It must be solvable somehow. ANY IDEAS?
Got it, thank you for your feedback. I will try to solve it in the next version.
@vjleoliu No worries. There are of course, other ways (extra steps) to address it. So I wouldn't compromise on the main purpose of the Lora. Thanks for sharing your efforts!
@Boobooberry In my tests, you can try combining it with another LoRA of mine. It can not only make the images more realistic but also improve the characters' facial features to a certain extent.
Amazing!
What happens with your PATREON?
The previous one has been deactivated. Please elaborate more here.
https://civitai.com/articles/22072/it-seems-that-ive-been-framed
Just wanted to note that in your examples, the Base Version images are by far the better ones. The plus version images seem like they've been in the oven a tad too long.
Going to see how it handles DAZ 3D renders, cheers!
I'm glad you like the Base version.
Hey, any flux kontext version of this?
In my shortlist, kontext is included, but obviously Qwen won out.
Absolutely fantastic LoRA! Keep up the great work!
Well Those conversions are easy, what actually does matter how well it transforms facial expressions or uncommon poses or body types
Is the trigger phrase actually 'changed the image into realistic photo,' or was there a mistake? In proper English, it should be 'change the image into a realistic photo,' without the 'd' at the end of 'change.'
Oh, thank you for your correction. English is not my native language, and I think I will fix it in the next version.
Where did you took that trigger to start from, it's no where on the lora page.
Where's the workflow? can't find it anywhere
I was asking me the same question as my result is far from what we see in examples. I may do something wrong in my workflow so i wanted to see 1 that worked to start from.
The best in the category...by far. But need to improve the hair, most of it seems a wig.
Thanks, buddy. I'll try to improve it in the next version.
效果真棒！大佬大佬！可以制作一个wan2.2版本的卡通转写实嘛，或者可以分享分享数据集嘛
天呢，我是国人的事情已经路人皆知了吗？最近跟我说中文的人越来越多，我是回到国内了吗？
wan22的写实还不错吧，有必要训练新lora吗？
"link to my patreon" is this the trigger phrase?
No matter what do you want, thank you for your attention.
Good model. Incredible write up and experiments.
Thank you for your recognition. If possible, please send some of your test pictures so that everyone can feel your "Nano-banana moment"
prompt: change the image into a realistic photo
maybe even without "realistic" because for klein it means pimples and freckles
Thank you for sharing. In my tests, Klein's "realistic" indeed has quite a few issues, and this is one of the motivations for me to work hard on making this A2R well.
@vjleoliu So far skin from this lora is reallly good after hirefix (normally klein tries "too hard" to make it realistic)
@Luntrix After hundreds of rounds of comparative tests, my feeling is similar to yours. It seems that Klein has magnified a small piece of skin, which makes its pores and freckles look numerous and dense, giving me a very uncomfortable feeling. You can imagine how I felt after looking at hundreds of such images; I feel like I'm about to get trypophobia. LOL