source quality
do you use the fp8 that the original model author publishes as source or do you do it from bf16
its the original from author, tried the bf16 and also fp32 bit it doesnt work as i havent enough ram and my gpu didtn even support bf16 so it upcasts to fp32.
i guess i need at least 72gb ram to do this, got 32gb and a very small ssd and slow hdd.
another problem i encountered (left my sytem running with hdd swap) is that the lora weight differ on fp8 to fp32 so even if i got it build the resullting images were totally different.
maybe there is a formula like applying weights/4 but i gave up on too much time consumed.
another point is that a whole bunch of loras is especially trained on the fp8 model, next problems...
its fine, its good, ill leave it as is and the image results of what i actually use Q5_K_M is nearly exact the same as the FP8.
waiting for more user input to V8 to release it.
I have the "recipe" for v4 , meaning the loras he used for that version of this merge at least, and I am periodically checking the loras and updating it for myself , I build it on bf16 edit-2509 model + loras then convert it to gguf and finally quantize it to q8_0, uploading my latest build to hf , going to share it here please compare when I do. (merging 6+ loras with bf16 qwen-edit-image really needs ram, I also have only 32gb ram, when merging this model with comfyui, I specifically make my swap file on windows 120gb , which helps , anything lower and it gives up.)