Ran training images on a skin-care product bottle of mine, but did not learn writing on label
Hi - :)
Would appreciate input on how I can get it to learn the writing on the label of a sunscreen bottle for example. Also, any way to train the models faster, currently taking me about 1.5 hours.
Hello!
As far as I know, Stable Diffusion doesn't know how to render text properly. Any model that uses an out-of-the-shelf, pretrained CLIP text encoder (like Dall-e 2), is not able to do such task. Other models like Imagen or Parti can render text, but they are not publicly available.
On the other hand, those are really big modes, so, normally, they take a lot of time to train, even if you freeze some layers and parameters. However, I have read that Dreambroth (newer research conducted by Google), can train Stable Diffusion in 15min approx. There is a GitHub repo for using Dreambroth with Stable Diffusion you can check. If you are looking for better tutorials, I guess in the following days some great notebooks will appear!
I hope this helps you ;-)
Hey Gerard,
Thank you so much for providing that input, I appreciate it. I have a few other questions would it be possible for us to speak for 15-20 minutes? I can be reached at aditya.addy.bahl@gmail.com
Please let me know! :)
Nvidia just release their paper about their yet to be released mode that handles text in images the best https://youtu.be/NSFU1DSA1ak