Phr00t My friend, your v13 version do not keep identity and had a lots of flaws, after hours of back and forward I almost done to combine the perfect lora match with the base model
so in my case I fix the auto crop in "TextEncodeQwenImageEditPlus" and create inside this node "target_latent" for both positive and negative which both connected to ResolutionMaster and from RM to Ksampler. this solve the uncrop issue in generations. after I solved this issue I had to play with the loras weights on difficult scenarios and pay attention to all details that not lose structure and identify. in short, if you like to take this bake model from me and use it feel free to ask, I am now in the process of backing
Agreed, any significant changes create a total identity change. Last i used 11.1 which still seems better at preserving identity
I've had good luck with character consistency with v13, but your mileage may vary and as I've said many times: use whatever version you prefer. There are tons of factors at play when making a model merge, character consistency just one of them. Qwen Image Edit themselves have been working to improve character consistency, which is suppose to be a key feature of their next release:
https://www.reddit.com/r/StableDiffusion/comments/1p3xlh4/qwen_image_edit_2511_coming_next_week/
Unfortunately it appears to be delayed, so we are doing the best we can with what we've got.