Dreambooth in Automatic1111. Cpu only & gpu option.

Key Takeaways

so great news everyone Dreamworld is officially available for automatic 11 11. so I thought we're gonna install that check it out see how it works and uh if you loaded your stable diffusion there's a tab here called extensions then you can go into available here and press load from now you have all the available extensions here you have dream Booth we're gonna just push install that has now been installed so we now need to completely restart the web UI now as you're starting this you're gonna get some more installations done because dreambooth is going to be automatically installed for you while this is installing I got something for you A friend of mine used to install kitchen work surfaces but uh he was arrested for counterfeiting anyway this has completed so we can reload our stable fusion and now we have the dreambooth tab here now you will still require a lot of GPU power for Windows it's about 10 to 12 gigabytes of vram but at least it's available to a lot more people now and it's actually actually now this is fantastic news for many uh many out there you can run this full 100 CPU [Music] now it will take a lot of time like overnighters many hours but it will actually work so you can do this on a potato but again it will take time we're gonna input our name here let's use Sebastian dream booth we'll get our source checkpoint is the stability Fusion model and then you can click create so now we're creating the model but we have yet to train it so that's gonna be what we're doing in the next step all right so we have a working director for this I'm gonna go into here the train model now there's a lot of settings here you don't need to touch all of them but some we need to touch the instance prompt is the prompt that you're going to use for your model so I'm going to use my face so I'm just going to put myself in there make sure that you put something in that that is not used in the training data already you can't put like man in there that's not gonna work you can put stuff like uh nonsense words stuff like that now you can use a class prompt like for example I would I would could use man here or if you're putting uh images of of dogs like your dog you could put dog here I'm gonna skip this for now but for fine tuning the model you can play around with the class images here's the path to your input images this is the folder of my images you can just go in here click that copy and put that right in there if you're using a class prompt you need to use classification images we were not using that here the training steps the good Baseline is the 100 steps for each image I think I have what was it 12 images I'm gonna do 1200 steps one batch size it's going to leave the learning rate Default Resolution is 512 is good if you want to save various models for various training steps you can input that here so let's say you have a training step let's say you have 20 images you have training steps of 2000 you could set this to or leave it to default at 500 that would mean you would get a model for training steps 500 a thousand fifteen hundred two thousand you can check out different models and the training steps and how they interact with your input images to fine tune what what you're looking for I'm going to disable this because I don't I don't want to do that let's check the advanced here yeah if you're if you don't have a GPU that can handle this you can use CPU only it's going to be as again as stated super slow we're gonna leave don't cash Latin so it's going to be a little slower but it's going to save on vram I'll train text encoder also uses a little vram but it's not required but you're you're gonna get better images if you keep that so if you can try to if you're running out of vram you can uncheck this again to say vram you can check 8-bit atom Center crop here is if your images are not cropped to 512 by 512 it's gonna automatically crop them I'd prefer if you just crop them beforehand make sure your images are proper and in 512 by 512 great in checkpointing we're going to leave that checked in and that's going to save us some vram but it's gonna run a little slower and a mixed Precision here we're going to set to fp16 to lower the vram even more and the rest here are more advanced settings so just leave them as default for now unless you really know what you're doing all right then we're going to start training now we got Creator select a model first so we forgot to select our model up here so this would be the one we created earlier and then click train so I wasn't able to record the training as the my computer was maxing out but once the training is finished you can choose to model in your models list I would recommend to use the GPU with at least 12 gigabytes of vram now 10 is technically possible it's a hard stretch as of right now it's brand new it's buggy but uh but it's here to stay Dreamboat in automatic 1111 really is the way forward uh because of its ease of use and easy install now if you can't get it to work there are still dream build collab options I would recommend both shivams and the last damn they're both great and whenever there's a new update or a lower vram requirement I'll try to make a new video out of it so yeah have a good one see ya