Can we continue to maintain it based on version 5.3?
If it's just for editing images, the latest 8.0 is not as good as 5.x. The most important thing is that it is much worse at maintaining the consistency of the original image. With the help of AI, I tried various prompts, but it was still very poor in the following aspects: 1. Facial features are almost impossible to maintain consistency; 2. Facial skin looks very fake; 3. It is difficult to completely maintain the background and the person.
Using 5.3 and i guess i stay there until the consistenfy issues are fixed.
Nobody forces us to use a specific version anyway
I hope @Phi00t update and refine the Loras base on 5.3 version.
Unfortunately the problem seems to remain at v9. I'm still checking every time there's a new version but keep v5.
Please, @Phr00t , consider most people want to use it for editing images! The text-to-image, anime or other such capabilities are secondary to having consistency. All versions of v9 seem to turn skin very artificial where on v5 this was not the case.
As a way to gauge this, just ask it to change something about the image. If large changes happen and the subject looks reimagined from scratch, it's not a good thing!
agree still using 5.3
一样在使用5.3版本，自从v7融合了“qwen image edit meitu”后，图像编辑出来的照片跟原图相比差太远，被强制上了一层美图效果。
Version 9.0 is better for consistency. give it a try
Version 9.0 is better for consistency. give it a try
I prefer the latest version (v9, not Lite)
After using V9, I tried going back and testing 5.3 again because everyone keep talking about it. I now agree 5.3 is way better than anything that came after it. I would pick up from there.
@TheNecr0mancer
Maybe someone will make their own merge using the recipe to make something better than v5.3? v9 works as intended
Version 9.0 is better for consistency. give it a try
When multiple people are involved or the background changes, the consistency of V9 is affected; it is no longer the same person.The V5 series, on the other hand, maintains good consistency.
Version 9.0 is better for consistency. give it a try
When multiple people are involved or the background changes, the consistency of V9 is affected; it is no longer the same person.The V5 series, on the other hand, maintains good consistency.
I think I might be noticing this too. Single character consistency looks pretty good in v9, but multiple characters might have something going on strange...
@Phr00t , multiple characters has ALWAYS been a struggle with Qwen in general (I mentioned this in my writeup from testing in v9). I don't think there is any way around it unless they change something fundamental in Qwen itself to deal with multiple characters better. What I had found that SOMETIMES works is if you have less detailed backgrounds. Qwen seems to distort characters when there is a lot of OTHER stuff to render. I was getting a bit better consistency with just the default Qwen 2509 workflow and just adding the loras I use, but its a lot of balancing and testing loras to make what I want a reality. Anyone generating for realism with your own images will always struggle using Qwen its just the reality of the tech for now until something better comes along. But for putting 2 characters in the same scene together, Qwen Image Edit is really unmatched. If they add better consistency for characters in the future, nothing will ever match it. Its really good at what it does, just needs refinement
thats why i most time remove background from all characters/things (qwen prompt or rembg), let qwen focusing on just the person details. last step either put two images together or jus say what background you want.
I also agree that v5.3 is still the best for consistency.