q3 bad
I tried q3-ks, very bad result
Can you share the command or workflow for how you ran? I’ve noticed it’s sensitive to hyper parameters and prompts.
But also it needs a different vae from other qwens, and the text encoder + vl needs to be passed. Do other quant levels work for you?
comfyui workflow, i used other proj's gguf, not that bad,of couse not good . The other proj's q4km works on some font image, but person pictures all failed,very bad
I tried q3-ks, very bad result
Diffusion models are very sensitive to quantization, you may need to use higher quantization levels unfortunately
Could you please upload the Imatrix file?
Maybe the value distribution is very different from LLM?
what is Imatrix? sorry i don't know that. that picture meams your method is not right .other project's gguf at least can work, yours can't
It was a question to
@danielhanchen
.
My guess is that no imatrix was used here because llama.cpp does not support imatrix for image generation yet.
may be not, the other project 's gguf can work.but i think your dynamic 2.0 method may be better than theirs