search

hatchetProject / QuEST

QuEST: Efficient Finetuning for Low-bit Diffusion Models
29 stars 2 forks source link

Simulated quantization or actual quantization? #15

Closed YangYang-DLUT closed 3 months ago

YangYang-DLUT commented 3 months ago

The current implementation of Q-diffusion is solely simulated quantization and doesn't have real benefits on speedup or memory savings. Your work outperform it, and if your method can act the actual quantization, it will be great progress. Could you offer me some information about this?

hatchetProject commented 3 months ago

Thanks for your interest! Our work (as well as most diffusion model quantization researches) also does simulated quantization. Actual quantization would involve much more sophisticated implementations and practical considerations. Unfortunately, we are currently not planning to move onto actual quantization. However, you can use utilize the quantization parameters to save the integer representations of weights for practical usages.

YangYang-DLUT commented 3 months ago

Thanks for your reply, still great work! 👍 About the Clip Score in experiment part, I am wondering about which tool or code repo you guys specifically used to measure it. Could you please offer me some information?

hatchetProject commented 3 months ago

Hi, the CLIP score is calculated based on the COCO2014 prompts. You can refer to this link for the calculation process :)