editing-images/leditsplusplus · Reproducing TedBench++ images

Hi, i'm working on an immunization algorithm to prevent images to be edited using diffusion models. I'm trying to use ledits++ as target method to defend from, since is more effective than other tecniques, like SEDit.
Until now, I'm struggling in reproducing correctly the images in your dataset. I'm using the correct parameters and the Ledits++ pipeline available on diffusers library for SDXL, but i'm not able to reproduce the same edited images. I use a generator with the given seed and i follow the step shown on the doc string of the diffusers code, but it doesn't work.

I've checked also your implementation: https://github.com/ml-research/ledits_pp, since as it's stated in the readme file the diffusers one in less correct because doesn't implement perfect inversion. But, still, the results of the pipeline are not the same.

I hope you can help me on this, and therefore, teach me how properly replicate your results.