[Request] Code for calibration dataset generation and Generic code for any diffusion model

[Ali-Flt]

Hi,

I wanted to use your work for comparison in my own research. But I couldn't find a straightforward way of generating outputs for a custom DM and a custom dataset of images.

Firstly, if I understood correctly, you have not provided any code for generating a dataset based on the algorithm described in the paper. You just asked us to do it ourselves based on the paper. Can you please provide the code for calibration data generation?

Secondly, I only found the code for calibration+quantization and image generation for the DDIM, LDM, and SDM models. What if someone wants to use your method on other DMs like DiT? Can you please provide a guide for employing your code in a more generic way?

Thanks!

[IC-Yuan]

Taking LDM as an example, it uses the def ddim_sampling function from /q-diffusion/ldm/models/diffusion/ddim.py for sampling. At the end of def ddim_sampling, it provides intermediates, which are the intermediate results at each timestep—these are the calibration set we need. Afterward, you can find the logs in def run from sample_diffusion_ldm.py and locate the saved intermediate results by checking for key == "intermediates".