I asked for just the white shirt in the photo to be changed to navy, and even the light coming through the window changed.
This time I added a puppy to a portrait, and the puppy looks fine but the background and the person look like a composite photo where they don't belong together. Anyone who has tried AI image editing will find this moment familiar.
It changes the spot you asked it to change, but it also changes the places you told it not to touch... It's a problem that's harder to improve than you'd think.
✦ The Two Models at a Glance
FLUX 3 Image Klein and Qwen-Image-2.1
FLUX 3 Image Klein and Qwen Image 2.1, both released recently, show this problem in different ways. Both support image generation and editing, and both accept up to 10 reference images.
But a comparison that tested them on the same images showed an interesting difference. Qwen was ahead at protecting the rest of the original, while FLUX was ahead at blending multiple elements naturally into one scene.
What Happens When You Merge Ten Images Into One Scene
✦ Experiment Order
How the references grew
A comparison experiment published on October 17 started from the same portrait. First it put a jacket on the man, hung a camera around his neck, and sat a Shiba Inu at his feet.
Then it added elements one by one, such as an arcade shopping street background, a bicycle, and coffee. It increased the reference images from one to ten, feeding the same images and prompt to both models. It also tried three seeds per model.
For FLUX, it compared separately the default setting with the prompt helper on and the case with the helper off and the original sentence entered as is.
✦ Multi-Reference Results
Same images, same prompt
The result wasn't a simple win or loss. Both models kept the person's face and didn't leave out the reference props. Even when ten images were used, the bicycle and coffee didn't disappear.
On the ability to include every needed element alone, both did their job.
The difference showed up at the moment the background and the person blend together.
With up to four reference images, the results didn't differ much, but once it went past five, some of Qwen Image 2.1's results looked as if a person and a dog had been placed in front of a shopping street photo.
The person looked large against the sense of distance in the background, and the dog lacked a shadow matching the surrounding light. It felt like new elements laid over a background photo left as it was.
FLUX 3 Image Klein's results were different. While keeping the character of the shopping street as a place, it matched the camera viewpoint to the person's height and readjusted the light and the shadows on the ground.
It made the person and the dog look as if they were shot from the same distance. In this experiment, when there were five or more reference images, FLUX 3 Image Klein got a better rating on the item "Did it blend naturally like a single photo?"
Both succeeded in bringing the needed objects, but FLUX 3 Image Klein was better at placing those objects in the same world.
The Opposite Result When Only Changing a Shirt
✦ Three Partial Edits
Both did the requested edit itself
This time they edited only part of the original photo: changing a white T-shirt to navy, erasing only the glasses, and adding the referenced Shiba Inu at his feet. In all three tasks, both models performed the requested change itself well.
The stitching pattern of the navy shirt remained, the glasses were gone, and the dog went into the specified spot.
✦ How Much of the Untouched Area Survived
Average PSNR outside the edit; higher means closer to the original
The real difference was in the parts nobody asked about. Qwen Image 2.1 kept the background around the shirt, the face, the lighting, and more almost exactly as they were. Even flipping back and forth with the original, it was hard to find what changed.
With FLUX 3 Image Klein, after the edit, parts that weren't requested, such as the background's brightness or bluish tint, the thickness of the glasses frame, and the brightness of the teeth, changed a little too.
So Qwen Image 2.1 showed strength in edits that change the spot to be fixed and lock the rest, and FLUX 3 Image Klein had an advantage in readjusting the relationships across the whole scene.
Touching up the overall lighting and perspective together is helpful for blending a background and a person naturally, but for someone who just wants to change a shirt, that touch becomes an unwelcome guest.
FLUX's Box Feature Is Quite Useful
✦ Region Selection Examples
A box is separate from preserving surrounding pixels
FLUX 3 Image Klein also has a feature for specifying the location to edit with a rectangle. Instead of vaguely describing "turn only the left shoe red," you wrap that shoe in a box and add the description. People who have used GPT Image 2.5 will know.
The feature where you tap a specific part of a generated image, leave a "comment," and send it. That's the one. In the comparison experiment, when only the left shoe was specified with a box and told to turn just one side red, only that shoe changed.
When the box was moved to the right, the right shoe changed, and when "just one side" was said without a box, both shoes turned red.
This feature is useful for placing an object at a specific spot on the screen or for telling more concretely where to edit. However, drawing a box doesn't mean the pixels outside the box are preserved as the original.
A feature that specifies the edit location and a feature that preserves the surrounding area are different problems.
And... honestly, from my own use of Flux 3, this feature of drawing a rectangle and editing only part of the image didn't work all that perfectly. But the editing of GPT Image 2.5, called the best image generation model today, isn't all that great either.
Anyone who has used GPT Image 2.5 has probably had the experience of GPT pushing the image only in the direction it learned, no matter how much I demanded changes.
It's extremely annoying, and no matter how angry I got at GPT or how I told it to improve, it never improved. Even when GPT passed along the right prompt, GPT Image 2.5 kept putting out already-learned outputs, as if it were a distilled model.
For example, I want a person to turn their body about 15 degrees to look to the side, but it only keeps turning the head without turning the direction of the body. However much I revised and revised and kept instructing... in the end I just gave up.
(Even when I threatened, "ChatGPT!! If you mess up again I'll hack you and kill you!!", it didn't get solved.) Even so, in this respect Flux 3 was better than GPT Image 2.5. If you keep trying, then when you luck into a good seed, the scene I want comes out.
GPT Image 2.5 is... um... people who have used it probably know. The quality of making the first image is excellent, but it really doesn't listen in editing work. It feels like working with a stubborn genius artist.
Flux 3 Image Klein, on the other hand, felt like working with a poor painter who tries hard to draw for you even if the skill is a bit lower.
There's one more difference here. The FLUX API provided by Black Forest Labs (you know this one is paid, right?) includes a 'pixel restoration' feature that preserves unedited areas pixel by pixel.
But this local ComfyUI comparison used the default template, and that restoration feature wasn't applied. So this result doesn't mean "FLUX changes the background however you use it."
It means that which execution path and features you use also decides the result.
Let's Look at Speed and Terms of Use
✦ Speed Comparison
Same machine, first run with ten reference images
Speed was also hard to sum up in a single line.
On the same machine, when feeding ten reference images for the first time to generate, FLUX with the prompt helper turned off took about 675 seconds, FLUX with the helper took about 804 seconds, and Qwen took about 1,038 seconds.
But there were also settings where Qwen got faster when generating several images under the same conditions after the reference images had already been processed. In other words, the order changes with the number of images and how you run it.
✦ License Comparison
Open weights are not unlimited use
If you want to download and use them yourself, you should check the licenses too. Qwen-Image-2.1's public license allows only non-commercial use for research and evaluation purposes, and you need separate permission to use it commercially.
FLUX also follows the FLUX Kommunity License, which has terms of use, so 'open weights' doesn't mean unlimited use. That said, FLUX's license terms are a bit more generous.
If you meet a few conditions, including annual revenue, commercial use is possible (check the FLUX Kommunity License (https://bfl.ai/blog/flux-3-image-klein)).
Let's Wrap Up
If you want to change only the shirt color or erase only the glasses while touching the composition and background of the original image as little as possible, Qwen is the better fit.
Especially when you already have a carefully made background and don't want to shake the person's face or the camera composition even a little, the ability to protect the unedited area is a big advantage.
Conversely, if you want to take a person and props from different photos and bind them into one scene, FLUX is worth trying first.
The more elements there are, the more the result depends on whether the person's size, the background's perspective, and the light and shadows blend together.
The difference between a feeling of reference photos simply placed on one canvas and a feeling of having been shot together at that place is bigger than you'd think.
But this is only an experiment result. Honestly, from my own personal use, Qwen Image 2.1 was better than FLUX 3 Image Klein across the board.
Qwen sometimes gives a feeling that people's faces turn a bit Chinese (?), but in my experience, even when combining several reference images, Qwen Image 2.1 was actually better.
FLUX 3 Image Klein, how should I put it, gave results that were so uneven that it was really hard to use consistently. In that respect, I felt that Qwen produced more stably consistent images.
Well, this is just my personal opinion, so if you like, I recommend trying both models!
That's all for this post. See you next time.
Related posts
- Nano Banana 2.1: 4K Images at $0.113 Each, Worth It?
- Amazon Bee Frontier and Eclipse: AI That Remembers Your Day
🌐 Korean version: 원본을 지키는 AI, 이미지들을 자연스럽게 연결하는 AI… FLUX 3 Image Klein과 Qwen Image 2.1의 차이
No comments:
No comments yet. Want to leave the first one?
Post a Comment