Turns out Qwen Image 2.1 isn't as broken as it seems. If you use the right settings and train a lora to fix some broken layers, you can get SOTA results from it.
This version is my first attempt. There are some things it still doesn't do well, but I'm adding to the dataset as I find problems.
The images in the right were generated with different settings and a sampler from RES4LYF. Get the workflow with the proper settings here.
Description
FAQ
Comments (18)
Thats impressive. Thank you!
Hmm...
ERROR lora diffusion_model.transformer_blocks.21.img_mlp.out.weight The size of tensor a (12288) must match the size of tensor b (4096) at non-singleton dimension 1
I'll push an update momentarily.
Try downloading again and let me know if it works for you.
It works. Thank you!
It worked for me? Its DoRA.
What was fix?
Dunno, whatever I downloaded clearly resulted in much better pics.
@Mescalamba ai-toolkit outputs doras in a slightly different layer format, so I patched it and fixed the lora.
what is it fixing?
The terrible, messy details.
@_Envy_ Which side is the fixed one?
@SolidCode You can't tell which side has terrible, messy details?
I can understand SolidCode's point, in many (almost all) cases the original is better composition wise but in some cases this lora helps to correct broken anatomy
@Kurovai Exactly. The original image looks better pretty much every time. With the prompts not included it also makes it impossible to tell which is more accurate or how poorly they were prompted to begin with.
I'd recommend labeling the lora examples and include the prompts to avoid future confusion, but personally I feel it ruins the aesthetics for limited gain.
@_Envy_what's exactly wrong with qwen, the only fix i needed was degrid so the qwen vae artifacts get mitigated
nice, sample images indeed look much better
It can make the image cleaner and more armature photo looking, but seems to hurt photorealism a bit with some prompts.
More of a tradeoff Lora than a Fix lora. You fix one thing but break another
I traded a hard problem for an easy one. I'm working on fixing the training data.