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!
FAQ
Comments (52)
I have to ask. Qwen Image Edit can already convert to a realistic photo natively, so what does this lora add to that? When I tried the prompt with and without this lora the results were fairly similar, though with this lora, the facial expression changed.
I don't know how you did it, but I didn't get very good results with my edits, so I trained it.
Moreover, although I hate to say this, Qwen-image-edit is indeed a bit unstable at times, and this LoRA can well make up for that.
These LoRAs are really trying to reliably trigger a built-in model function, but as you say a pretty obvious prompt usually gets the base model to do this anyway. However, with both Qwen and Kontext, I've found other 'style' tasks that do not usually trigger on certain types of input image. So there are occasions where a trigger LoRA is useful.
I found that just putting 'make this image real' consistenly did a great job without any lora. I took images from various 80's/90's cartoons and the results were great. Maybe about 3/50 didn't like quite right but the rest wrre spot on.
@MeditatingHamster I have revised the model introduction and added the effect of using only prompts to the comparison chart, making it convenient for everyone to conduct a visual comparison.
@MeditatingHamster From my own experience the native outputs looked great but with human females, I very quickly noticed a lot of same face. This happened whatever the source image was. I'm hoping this can solve that.
I’ve been waiting for this sort of thing since I first used img2img. Qwen can sort-of do this, but I’m looking forward to good results with this.
Looking forward to your feedback
Would you mind if I uploaded this to tensor for online gen? Given that the only way to use LoRas with qwen-image-edit is using the workflow feature, I’ll probably be the only one using it, but it would be more convenient for me.
Or rather the other way around: Would you mind uploading it to tensor? As the original creator, you’d get some online credits from people using it, if that interests you. Either option would be nice for me :)
@aiaiaiai829569 I'm looking for an international platform that can be used online. Is Tensor quite popular?
@vjleoliu tensor is one of the biggest ones, afaik. For some reason it does not, currently, allow loras for Qwen-edit in the basic interface, but you can use them in workflows and the online "tools" section.
If you are looking for profit, it does allow setting tools to require a fee before use, or a fee for download, or some other options. Not sure if tensor is the best for that, though. I have not researched best sites for making money with tools.
flux kontext of this lora would be awesome, qwen image edit is too heavy on my laptop 😥
Everything will get better gradually. Think about when FLUX was first released; didn't it also have very high requirements for computer configurations? Let me tell you secretly, nunchaku's Qwen-image has been launched, which can generate high-definition images in tens of seconds, and Qwen-Edit won't be long in coming either.
@vjleoliu What more optimizations could be done on QWEN EDIT?
I think nunchaku is the limit.
@zerocool22 Improve the generation speed. Your response is very timely. Today, nunchaku's Qwen-edit has just been released.
@vjleoliu Nunchaku, is bettere than GGUF Q8 or FP8 models, i mean, Nunchaku wins on speed generation and less loss quality?
PD: I ask because i'm thinking on trying Nunchaku for first time, since i hear a lot about it.
@zerocool22 Yes, your guess is correct.
please add a workflow for comfiui
yes, i did it
i copy your nodes and slot in all model and how come i use your nodes, but outcome not the same as the reference image?
This lora kicks ass!
Glad to hear you say that, thanks buddy
I don't understand why does it LOAD AGAIN the wanVAE every prompt? I just changed a prompt and it loaded again :/ and it takes me a lot of time, look:
Dude, please next time make SAVE nodes instead of PREVIEW, I have been running the workflow thinking it was saved in my output images! And I just discovered it was just previews :/
I'm really sorry, man. I'm used to previewing multiple times and then choosing a satisfactory image to save.
But I can tell you a little trick that might help recover your images. You can find all the preview images in comfyui/temp. You can try that next time.
Remember, if you restart comfyui, the images in temp will disappear.
EDIT: I commented on the wrong thing. I meant to comment on a workflow that said to use this lora. Oops.
Original message:
I ran into some issues. FYI, the FusionX Lora is locked inside the Wan Loader SubGraph, meaning the user will get errors if it isn't that exact name and path, and won't be able to find the place to fix it.
FusionX Lora？but this is a lora of Qwen-edit, so……do i miss something?
@vjleoliu Ah shit. I'm sorry. I commented on the lora. I meant to comment on the Anime 2 Real workflow that suggested this lora. Sorry about that. But yeah, it strangely uses Qwen and Wan together.
Which workflow? I didn't use "wan" in my workflow.
@vjleoliu It was this one. I just plain misremembered the suggestion to use your LoRa. Haha sorry for the confusion. But using Wan is a really wild choice.
https://civitai.com/models/1882882/anime2reality-qwen
@Jellai I took a look at that workflow, and it seems that it only uses Lighting's LoRA, not mine. Also, it's true that some people use wan to generate realistic images, and I sometimes do that too.
so I see that it works just with the prompt: turn this person into a realistic photo" or something like that. In the samples you prompted the image itself like: "girl in jumpsuit holding this and that..." is this necessary? will this make it better?
Because the scene was described, the characteristics of the character were also described incidentally. Of course, this helps the model understand the character better.
Hi i have tried using qwen edit 2509 and the usual qwen but i was not able to do it like urs, not too sure what has went wrong
I don't quite understand what you mean. You can send me the image you want to convert, and I'll try to help you with it.
@vjleoliu oh how should i send it to you?
@leonviz ummmm，Join my Patreon channel and send it to me in the chat room.
Some of these comments are comedy gold!
I also have an issue and require immediate help!!!
Why?!?!? when I click the small traingle (next to where it says "Queue") does a single image appear in the box and not a 3 hour blockbuster movie directed by Michael Bay?
I asked the computer for a movie ?????
lol, this is the first comment on Civit that had me looking to see if there was a way to send buzz to a comment.
Did you feel attacked? Why so salty? LMAO
@imonaboat1 You actually understood it?
@vjleoliu ..yes? it's parodying other comments expecting miracles, and quite well.
@imonaboat1 Oh, haha, when you put it that way, I just realized it. I didn't get it before.