runpodctl VS EveryDream2trainer

You can rent a GPU on https://www.runpod.io/, which also has stable diffusion templates so any time you start the GPU SD will be preinstalled for you, ready to use:)

Another option is to use a service like www.runpod.io to rent time on more powerful systems. A few times a week I’ll load up whatever the latest 13B or 20/24B (and even low bpw 70Bs with EXL2) model is on a system with an rtx 3090 for $0.44/hr, and sometimes I’ll treat myself to an A6000 system to run 4bit 70Bs for $0.79/hr. They also offer A4000 systems with 16GB VRAM which is plenty to run a 4bpw 13B EXL2 model or an 8bpw 7B model, and those systems are just $0.36/hr.

Many people use sites like runpod for this.

You can rent time on systems on www.runpod.io with a 48GB A6000 for $0.50/hr spot pricing and $0.79/hr regular pricing. 3090s can be had for $0.29/hr spot pricing and $.44/hr regular.

Keep in mind you can also use www.runpod.io to rent access to systems with a 3090 for about $.45/hr. It might be significantly cheaper ir at least more affordable to do this for a few hours a week instead of dropping $1,000 on a new laptop. This is what I personally do (I generally use it in the evening, and can get an Nvidia A6000 with 48GB VRAM for $.49/hr spot pricing). This lets me play with the latest 33B and 65B models, with really fast replies, and I spend maybe $5-7 a week if I use it a lot.

you can also connect the volume to other cloud storage (pcloud, dropbox, etc) (look for the option cloud sync in the "my pods" console) , or use python libraries to pull/push files from S3 buckets / dropbox / ftps / your NAS / etc, or use their docker tools to transfer files between your local PC and the volumehttps://github.com/runpod/runpodctl/blob/main/README.md

I'm using EveryDream2 with SD v1 based models. You can define whatever resolution you want for training, as long as your Vram allows it.

There is a demo optimizers settings here that uses special settings for the text encoder: https://github.com/victorchall/EveryDream2trainer/blob/main/optimizerSD21.json

If you have a large dataset of pictures, I'd recommend https://github.com/victorchall/EveryDream2trainer instead of Dreambooth. It has decent documentation (well for an open source project that is) and it has a very nice validation feature (disabled by default) which actually gives you good feedback on how the training is progressing.

EveryDream2 can handle this. They even have tools to help you autocaption using BLIP. https://github.com/victorchall/EveryDream2trainer

I recoment EveryDream2 for training, it has a lot of nice features. I'm not sure there is a proper manual to learn how to train, but there is a lot of information available. I have been learning this subjects for a few months myself.