Translate: Q3 isn't very good > “q3 不大好”
估计还得手动调权重吧，就像edit一样，q4km还行，以下的估计都没法用
Hi, I’d appreciate it if we could chat in English.
It’s possible that the precision needs to be increased. So far, the quantization being used is the same as the one used for Qwen Image Edit 2509 when we updated the quantizations.
Are you using it without the Lightning 4-step LoRA?
edit 2509 q3 is very good,but layered not, q4km is not good either,while q3 is very bad
I understand what you're saying. But I was asking if by any chance you were using the Lightning LoRA of Qwen Image Edit (also 2509) in your Qwen Image Layered generations?
@ZKong It would be great if you could create a table of the results you get using the same input image and seed across different quantizations, so we can compare them.
Please test the lower quants; if anyone else is encountering issues, we’d appreciate it if you could try using other quantization levels as well.
This is similar to what happened with Qwen Image Edit 2509, where multi-image support was added but lower quant outputs were poor before we updated. We are going to test different precision quantizations for the lower quants to find the right balance between keeping the model weight low and ensuring better outputs. Once we have the results, we will provide an update.
Did not use lora first, when using q3,q4 . q3 very bad,q4 better. Edit 2509 Lightning LoRA can cause bad results too
It is normal for LoRA Lightning to degrade results. The same thing happened with the Qwen Image Edit LoRA on Qwen Image Edit 2509...
Please send your comparison table. We should be updating the quantizations in the next days. 👍🏼
i deleted the q3 long ago😄，waiting for your perfect quantizations like edit 2509