Have you tried silveroxides/convert_to_quant ?
#3
by baesik - opened
It has preset model filters to prevent quantization of the best layers and you can run a learned svd quant. Official comfy method is --int8 --scaling_mode row --convrot --convrot-group-size 256. Excluding the --simple flag will run a learned quantization that looks a lot better.
Not a fan of these star_model_converter quantize models.
The referenced repo author mentioned the following in a discussion: https://huggingface.co/LAXMAYDAY/qwen-image-edit-2511-int4-comfy/discussions/1#6a4c8bdca6af4e156c4c40b9