Download the sampler used in the workflow here (or find it in the comfy mod registry):
Description
FAQ
Comments (4)
https://huggingface.co/madebyollin/texture-fix-vae-for-qwen-image-2.1
Use this too.
Qwen Image 2.1 is touchy when it comes to getting right sigmas, like really touchy. Sampler/Solver is important too, but sigmas even more.
Issue is that ComfyUI hardcoded one sigma and that aint right and for most optimal results you need original designed floating sigma that depends on image size (I would say image size AND longest side).
If you ask right AI with enough context, its able to fix it. I have solution but I have a lot of other stuff tied to it, so it wouldnt be that easy to publish.
Should add, I merged 2.0 into checkpoint, works nicely. I would say it follows prompt a bit better than 1.0.
@Mescalamba Nice! And yes, it should. 2.0 is retrained from the ground up and then, through trial and error, I filtered it down to just 3 of the 29 blocks, where the problem actually was. It leaves most of the weights alone.
And thanks for pointing out that VAE!
@_Envy_ I know only that filtering some late blocks (like very late blocks) for most LoRAs improves sharpness quite a bit, without much negative impact.. would need to check which parts exactly, think its gate_up and something. Author of head swap LoRA figured most, no clue how yours works tho, I guess they messed up something quite a bit when creating Qwen Image 2.1 ..
And ofc, thank you for making this LoRA!


