How low is acceptable for you? (the users)
Thanks everyone involved but I got a general question for people who use this:
What quantization do you prefer? and how much vram do you got?
I just tried the qwen-image-2512-Q2_K.gguf, and ended up with a black image? and it was already horribly slow?
Maybe the 8 vram I got in front of me does not cut it?, And I should just stick with z-image turbo (which i know is in no way comparable technically) for this card?
We've been working on some better Q2_K, Q3_K, and Q4_K versions. Should come out later today!
Hi just uploaded. In particular you should see a massive quality improvement in the Q2 and Q3 K versions!
I just tried the qwen-image-2512-Q2_K.gguf, and ended up with a black image? and it was already horribly slow?
@Kwissbeats If you have, remove the commandline arg for sageattn in the bat file, i.e. remove this --use-sage-attention.
This will take care of the black image, but not slowness. Additionally, you should enable live previews so that you can see the image as it is being generated.
" What quantization do you prefer? and how much vram do you got?"
q4_0 Qk_M, i have i7 13th gen rtx 4060 (8-vram) laptop 40gb ram 5600mt/s. works fine, arround 2-3 mins per gen, also used wan2.2 q4 and worked.
" I just tried the qwen-image-2512-Q2_K.gguf, and ended up with a black image? and it was already horribly slow?"
I tried the qwen2511 with no problems, same time with q4 of 2512, but the result is an extremely saturated image.
" Maybe the 8 vram I got in front of me does not cut it?, And I should just stick with z-image turbo (which i know is in no way comparable technically) for this card?"
-i set de comfy on low vram mode, modifying some file(portable version).
The vram mode shouldn't affect image quality. The 2511 and 2512 models are for different purposes. If you're following the example workflows and guide, you should not see a black image, but you may need to play around with you sampling steps or hyperparameters to get the best quality for your prompt. https://unsloth.ai/docs/models/qwen-image-2512