Fast Dream Booth - Updated Tutorial - April 2023

Learn how to use Fast Dream Booth to create custom tuned Stable Diffusion models. This in-depth tutorial covers all the basics of running The Last Ben Fast Stable Diffusion Google Collab notebook. If you have any questions please leave a comment below.

[GarethMcD-zqei]

Much appreciated! I've actually done that my self, the whole nightcafe thing, it's a good way of getting starter images like to use in img2img or even use your results from nightcafe in inpainting & it saves you having to wait on your own PC just to get some quick mock images to work with, it's not for everyone, but it could help as you clearly showed in the video.

[diggydretv]

This is great! Is there a way to do this to make a Lora?

[RobertMichael-jhfu]

What classifier images does this use? I don’t see an option to set up a custom set of reference images for the class ie person, man, woman, style ect..

[CRIMELAB]

as of recent, my ckpts not working when i upload them to tensorart & seaartai. i get generation error .
:((((

[DividenderoCL]

Your work is very good. I have a question, did you use Fast Dreambooth to create the "dvmech" model?

[OwaOwaBiowa]

how can we get more trigger words, u got just your image name but there should be a way toget trigger for for outfits to if know please help

[kasteh]

I have one question, i can I train a civitai model? Im trying to do make with "Adobi's Real Flexible Mix [aah!]", but always gives me an error. Could you make some video doing with civitai models? please.

[dissonancE..]

I'm training a model for a model (I'm the photographer). We gave hundreds of high quality images that can be used. I've found a dataset of around 75-80 to be the best, but I also noticed that I needed to use regularization images and increase the text encoder steps to get cleaner, less deformed, results. While if I increase unet steps @ the 1:75 ratio I find that it'll often get overtrained, so I stick to around 3.5-4k training steps. Would you say this is normal or should I change the learning rates? I've trained on base models: SD1.5, SD2.1, Dreamshaper w/ & w/o baked vae, & RealisticVision

[devoidsloth]

how would I retrain my model with this? I accidentally overtrained it.

[annyeong]

It would of been cool if you actually compared it from the base 1.5 (or any other model based on 1.5) and then showed the difference because they do actually do (rather inaccurately but still) recognize and create most popular artist faces.