search

thu-nics / MixDQ

[ECCV24] MixDQ: Memory-Efficient Few-Step Text-to-Image Diffusion Models with Metric-Decoupled Mixed Precision Quantization
https://a-suozhang.xyz/mixdq.github.io/
27 stars 3 forks source link

How reproduce the FID #13

Open suppwilly opened 4 weeks ago

suppwilly commented 4 weeks ago

Hi, I can't reproduce the FID of FP16 SDXL-Turbo in you paper (17.5), could you share the prompts you use and the method to crop the image, as you know, each picture has multiple prompts in coco2014 and it will affect the FID, the crop method affact the FID too.

A-suozhang commented 3 weeks ago

Sry for the late reply, we are organizing the eval script, and would release it soon. Please stay tuned for the code update.