Closed KohakuBlueleaf closed 11 months ago
yes, you can quantize your model with PTQD. It works fine with a higher bitwidth (>=4). If your model structure varies significantly from unet used in stable-diffusion, the code for reconstruction may need to be modified.
@ThisisBillhe Thx for informations!
I have other few questions
1, will you support SDXL in the future? which use different repository (the sgm from stability.ai) 2, will you consider to make/impl/support sd-webui or comfyui? which are common gui client for stable diffusion
I can try to do these 2 by myself but I want to check if you have any plans on it
well, I am currently working on the research of extremely low-bit quantization of diffusion models, so the answer is no for now.
@ThisisBillhe ok!
@ThisisBillhe Thx for informations!
I have other few questions
1, will you support SDXL in the future? which use different repository (the sgm from stability.ai) 2, will you consider to make/impl/support sd-webui or comfyui? which are common gui client for stable diffusion
I can try to do these 2 by myself but I want to check if you have any plans on it
Hi~Have you worked out with low bit SDXL? And would that be convenient for you to share some data about the quantized model?
As title If I want to quantize my own model can I just use the script in PTQD? or we need to change something like dataset for reconstruction or bits width