Is the result consistent with stable diffusion official when using default parameters?

[xpeng]

Since I used lexica.art or dream studio official for a while to compare result, so we always share seed number and other parameters with friends.

Is the result in InvokeAI consistent with stable diffusion official when using default parameters? or did i miss something?

Thanks!

[lstein]

You may or may not get the same images. The code bases started out the same, but have diverged over time. Your best chance for reproducibility is to make sure you are using the same sampler. Older versions of Stable Diffusion source codes defaulted to plms, but newer ones tend to use one of the k* samplers, such as k_euler_a. We use k_lms.

Macs, due to a difference in their random number generator library, do not produce the same images as Windows or Linux machines.

[xpeng]

You may or may not get the same images. The code bases started out the same, but have diverged over time. Your best chance for reproducibility is to make sure you are using the same sampler. Older versions of Stable Diffusion source codes defaulted to plms, but newer ones tend to use one of the k* samplers, such as k_euler_a. We use k_lms.

Macs, due to a difference in their random number generator library, do not produce the same images as Windows or Linux machines.

Thanks lstein, i'll keep trying.

I think there may be some random number generator different here, i found there are difference between using "cpu" and "cuda" device in manual_seed(seed) too.

[psychedelicious]

@xpeng , we have PR #981 pending, which will allow CUDA users to generate noise with CPU and at least have some consistency within our implementation.

Closing as the question has been answered.