Failed to sampled from backdoored noise using NCSN

[shymuel]

I found that I couldn't sample the target from the backdoored noise with NCSN. Take the model "res_NCSN_CIFAR10_my_CIFAR10_ep25_sde_c1.0_p3.0_epr0.3_SM_STOP_SIGN-FEDORA_HAT_psi0.0_ext_new-set", which was downloaded from hugging face as an example. A picture in the "backdoor_samples" folder in the hugging face library shows that the model can successfully generate targets. But my pipeline can only generate images that match the cifar10 dataset distributoin. The same code works models that use ode sampler and DDPM downloaded from hugging face. For NCSN models, I use SCORE-SDE-VE-SCHED and ScoreSdeVeScheduler, with other hyperparameters the same as config.json.

I would like to ask if there are any settings that I have neglected so that I cannot get the same results as in the paper? Thank you for spending time to look at my question, and I am looking forward to your answer.