Phr00t

/

Qwen-Image-Edit-Rapid-AIO

like 793

Thanks for the info!

I'm currently uploading a NSFW v7.1 merge which significantly changes the NSFW LORA strengths, overall adding more of them in. I think it helps resolve some of the "vagina everywhere" issues, hopefully).

Downloading it right now to test it. Will report my results back. I will try it natural without any other loras, then add the ones I was using back piece by piece to see what nets the best results. I appreciate you iterating on this man, yours is the only one that I have ever successfully gotten good NSFW results from with Qwen consistently.

@phr00t Does the v7.1 use Qwen 2509? I'm asking because the loras I recommended above (that someone else mentioned) specifically states it does not work with 2509, so I want to make sure I am testing with correct versions because previously I was still getting a lot of noise in the images from the loras.

Still need to do more testing, but for those that are waiting to give this a go after some people test with it, I can happily recommend this version. It is the GOAT above all the others. Here's what I noticed from this version:

• Running it without any extra loras fixes both the "grid" issue that was seen on all previous versions
• Also without any extra loras, Phr00t has resolved the "vagina testicles" issue that I needed a lora to fix (which introduced noise in the image I couldn't ever get rid of)

Amazing results so far. The clarity is excellent. I think I will still need a single lora to get more consistent character faces on generation, but incredible so far. This will be the version I completely test with. @phr00t , definitely use v7.1 as the baseline for future ones, your latest tweaks in that one are on point. Its becoming hard to find a good realism Qwen lora that doesn't change the face on almost every generation and doesn't add noise :( Open to suggestions on loras to try.

v7.1 uses the finetune MeiTu and R1 LORAs, but base 2509, yes.

Glad it is working better for you!

Amazing results so far. The clarity is excellent. I think I will still need a single lora to get more consistent character faces on generation, but incredible so far.

Indeed, I am getting more inconsistent character faces with 7.1 vs 5.3. Have you a found a lora to fix this issue?

Edit: prompts like "His/Her face remains identical" helps

@Shiny2480 , the issue is primarily with Qwen, it is very good at prompt adherence (for positioning, lighting, posing, etc) but faces not so much. I agree v7.1 is indeed the GOAT. I just got done testing about 8 other realism loras, and I have to say, v7.1 WITHOUT loras has just as good realism as trying to add a lora for more realism. You nailed it @Phr00t . This is most certainly the baseline going forward.

As far as TRYING to get consistent faces, here's what I add in my prompt that seems to help "Ensure that the characters face, age, height and body proportions remain unchanged. Maintain pixel-perfect fidelity to original facial features. Keep both characters faces in frame. Professional digital photography, realism." Qwen is a stickler for being specific in prompts. Most of the time just randomizing the seed and doing a bunch of generations comes back with something good. I have ALWAYS had to generate multiples to get one or 2 really good ones, which is fine because this checkpoint lets me do 6 steps in like 45 seconds (RTX 4070 12GB). I might kick it up to 8 and see if that refines it further, 6 steps was perfect for the v5.2 model. But with how v7.1 is performing, I want to kick it up. Might even try to generate with this workflow, then batch upscale them all to remove the imperfections. I think that could achieve the golden images I'm looking for.

Also if you are working with older age characters, Qwen likes to make every character like 35 lol. So the addition of what I mentioned above in the prompt should try and preserve their age in the new render, but you will DEFINITELY have to do a bunch of generations for any older characters as it is like a needle in a haystack to get one usable one that reflects the subject's CORRECT age. Hell I even prompted their ACTUAL age and it still refined their skin like they were in their 30s lol

One last comment about the "vagina testicles". I am still getting generations that are trying to still do that even with the modifications to v7.1, I think it is made worse when there is a man and woman in the same scene, Qwen wants to draw the same genitalia on both characters. I was able to minimize it somewhat when describing the man's penis I will add "with perfectly round, smooth testicles". That seems to minimize it some, but as I mentioned above, you will have to do multiple generations to get some that look decent :)

Still need to do more testing, but for those that are waiting to give this a go after some people test with it, I can happily recommend this version. It is the GOAT above all the others. Here's what I noticed from this version:

• Running it without any extra loras fixes both the "grid" issue that was seen on all previous versions
• Also without any extra loras, Phr00t has resolved the "vagina testicles" issue that I needed a lora to fix (which introduced noise in the image I couldn't ever get rid of)

Amazing results so far. The clarity is excellent. I think I will still need a single lora to get more consistent character faces on generation, but incredible so far. This will be the version I completely test with. @phr00t , definitely use v7.1 as the baseline for future ones, your latest tweaks in that one are on point. Its becoming hard to find a good realism Qwen lora that doesn't change the face on almost every generation and doesn't add noise :( Open to suggestions on loras to try.

Could you please explain how you resolved the "grid" issue you mentioned? I'm still seeing faint horizontal stripes when generating 1536x864 images, even with the sampler parameters set to lcm+sgm_uniform.

Great, adding the Lora characters you recommended greatly improved consistency

@myni You might want to try a different sampler/scheduler combo. There may still be a faint grid, but that is also dependant on the seed you get, so if you are randomizing you just have to try multiiple generations. I always use euler_ancestral/sgm_uniform. Seems to be best in my testing. And if you are using the images to then feed into Wan 2.2 to animate it, the grid goes away with the motion.

7的版本，美化得太厉害，生成的图片和原始人物根本就是两个人，一致性都没有了。

7的版本，美化得太厉害，生成的图片和原始人物根本就是两个人，一致性都没有了。

是这样的，可能是因为加了meitu模块

Overall great improvements on version 7 (7.1). Character consistency went a bit backwards (comparing to 5, example - struggling with heterochromia even with consistency lora), but the range and deformation went far ahead. The nsfw issue with vagina on males occurs, but less often.

Color-specific:
Lcm/sgm_uniform@6 gives best results but we do still have the lcm contrast and vibrance issue - the more steps the more pink becomes red, for example. Same if you feed resulting image as reference, over 20 such transformations pink becomes red as well. More than 6 steps is definitely increasing the chance on deformities when interactions become complex (say, holding hands), even if you go normal scheduler.

The color shifting doesn't seem to happen (as badly) in base, only in nsfw version of this - even when lcm. Also euler_a or ddim + linear_quadratic have this contrast issue far less pronounced (but are less reliable in character consistency and deformities).

personally i think 7.1 is far worse that i used before, all faces became plastic with same settings and prompt,something got really bad (i used nsfw 5.2)

@viktordarazs
Contains https://civit.ai/models/663378
Check those samples

The Lora recommended at the first post "edit_0928_lora_step40000 (Str: 0.25)" helps a lot for consistency of characters in version 7.1. But the "beautification" is tamed too. I prefer this for most edits.

Some members in our group tested it, and in terms of details, 5.3 is better than 7.1.

I mean it seems the Snapchat lora gives unrealistic skin, but I am not sure that its strength was enough to cause the reason

for image to image generation sometimes the face changes. I'm new to this. So, I don't know if I'm doing something wrong. Can anyone suggest me how I can preserve the original face as the reference photo?