Flux-Kontext Issues With Body Proportion
I am just experimenting with Kontext, and find it does some things very well, but it seems to really have a problem when using a source image that is just face and upper body and then making the end result be a full body image. Here are a couple of examples I made. The first image was made using the default workflow and prompt. The head is way too big for the body.
In the second image, I modified the default workflow to include two images. The issue a little better if I only create square output images, but I really would rather not be limited to that.
Is anyone else seeing this, and have you found a workaround?
I feel like this is a technical problem that is difficult to solve.
I have also encountered this situation.
I think we can try to solve the problem through lora.
However, this does not seem to be easy to solve. After all, the model author should also pay attention to this problem, but it was released without being solved.
I don't know if Pro is like this, and whether there is a downgrade.
Perhaps, you can add prompt words such as: keep the proportion, look good
some one should train decent loras doing this : 1- bigger head make head bigger , 2- smaller head make head smaller
I also noticed this problem, but it is related not only to the head, but also to the legs (they are too short) and the torso.