search

TheLastBen / fast-stable-diffusion

fast-stable-diffusion + DreamBooth
MIT License
7.55k stars 1.31k forks source link

Text encoder steps #2898

Open pravbk100 opened 3 months ago

pravbk100 commented 3 months ago

I have trained with 10/20/30 images and 500 images with good success. The only difference i can see with lesser dataset vs 500 dataset is - lesser one tends to be a bit limited and 500 one is bit flexible. For Both lesser and higher dataset, i kept the text encoder steps at 350. Ben mentions 200-450 is enough for small dataset. So what is the good text encoder step count for higher dataset like 400/500 images. Is there any calculation for that? And also whats the good step count for unet for that amount of images. Thanks

TheLastBen commented 3 months ago

For a larger dataset, increase the TE steps to 1000 or 1500

pravbk100 commented 3 months ago

Thank you. What about the ideal unet steps? Is there any formula?

TheLastBen commented 3 months ago

not really a formula, you should easily figure out the right settings once you get used to the model, keep trying until you get the hang of it