about rebalance lora.
Hello, I’ve noticed that the “rebalance LoRA” works very well, but when applied to v7.1 it introduces horizontal-band artifacts. I see you’ve already released v8—could you consider merging this LoRA into the next version?
v8 is meant to improve artifacts, so give it a try.
It sounds like "rebalance" is trained on cosplay data, which I'm afraid would tilt the results toward such results. The "base" Qwen Image Edit AIO should be as "general purpose" as possible. If there was more variety in the dataset, I'd consider it further.
v8 is meant to improve artifacts, so give it a try.
It sounds like "rebalance" is trained on cosplay data, which I'm afraid would tilt the results toward such results. The "base" Qwen Image Edit AIO should be as "general purpose" as possible. If there was more variety in the dataset, I'd consider it further.
Thank you for your help. How can I load this realism-enhancing LoRA without encountering such severe horizontal artifacts?
Currently, I'm using a two-stage sampling approach: a total of 8 steps with an LCM/Normal scheduler. I apply the LoRA for the first 6 steps and bypass it for the final 2. While this improves the results compared to using the base model directly, the output is still noticeably inferior
Are you using v8?
Are you using v8?
"Yes, I was previously using V8, but when I additionally mounted the Rebalance LoRA, it caused severe image degradation—mainly noticeable as very prominent horizontal banding and a significant loss of clarity. However, this issue did not occur when using the official Qwen Image model. I see that you have now released V9—thank you very much."