Extending Dreambooth with Kane Wallmann's implementation about diffusers

[ShivamShrirao]

Hey, I have had this in mind for some time, just haven't got enough time yet to complete it.

[1blackbar]

THat would be really great to have ! Id love to train multiple subjects at once instead of keeping new ckpt for each

[Hugo-Matias]

Thanks Shivam, that's great news! I don't have the required expertise to be much help other than spreading the word but it's nice to know you've already considered it.

Someone posted a similar implementation on reddit a few days ago, might be useful to have a look as well.

Since most people are training models on several different subjects/styles, having a way to train everything in a single model is desirable as managing multiple 2GB becomes wasteful and unpractical. This feels like the logical next step for Dreambooth.

[ShivamShrirao]

Added in 351f3b6

[Hugo-Matias]

What a legend! That was lightning fast, really eager to try it out. Thanks once again.

[Hugo-Matias]

Sorry for nagging, a couple of questions popped into mind when I was training with the new method.

Is the number of class images parameter set per concept or is it the sum of all? For instance if I have a concept with a "person" class and another with a "style" class. Setting num_class_images to 500 will it pick 250 or 500 from each path?

Can we continue training if instead of pointing to huggingface we use a diffusers model saved from a previous session? Is that feasible or will it mess the model regardless of the parameters you use?

[ShivamShrirao]

@Hugo-Matias per concept.
500 from each.

You can continue training.
Depends how much it was trained and on what.

[Hugo-Matias]

Awesome, thanks for the fast reply.

Sorry for taking too much your time, forgot to ask this observation from the latest training.

Are the concepts trained in sequence or the steps constantly switch between them?
Let's say if I hypothetically have 2 concepts of 10 instance images each, so an epoch is 10 steps. What would happen if I set a maximum step count of 6? Does it do 3 steps on each or just 6 on the first? I'm asking this because I trained 2 persons with the same class, same amount of instance images and it looks like the second one is a bit worse in terms of resemblance. Perhaps I should train it alone to see if there is something wrong with my numbers or the dataset used.