Support for multiple fine-tuned concepts & classes on one model with Dreambooth about diffusers

[ShivamShrirao]

@jslegers Yeah nice suggestion, I was also looking into it, Just haven't done enough testing yet to see how well it works.

Or I might just update it with an option first so that people can try it atleast without ruining the current results.

[carbocation]

Interesting - the @kanewallmann approach (permitting setting the prompt word and class word on a per-image basis) is very appealing. (Appealing as opposed to, e.g., training a second time on top of your already modified model, since additional retraining in that sequential fashion likely harms the model's memory for original concepts.) My barrier to using it seems to be the same as @jslegers: RAM limitations.

[IdiotSandwichTheThird]

@ShivamShrirao if you do add it, please make it use the caption for each image from a .txt with the same name, since it's a lot of easier creating that with automatic's blip captioning, rather than manually naming each image.

Also there is just a ton of issues with long filenames going the caption in filename way, it just seems really subpar.

[ShivamShrirao]

@IdiotSandwichTheThird It has been added now. 351f3b6

Yes, it makes use of a json to get names and class.