Suggestions on how to train a subject that have text?

[ergosumdre]

I'm trying to train the model on a series of images like this:

However, when I render the images after training, there is always an issue with the characters. The image either has missing or extra characters. For example:

Can anyone help suggest ways to improve the model?

[TheLastBen]

Text isn't something diffusion models are good at, even Dalle-2 isn't.

[giteeeeee]

Try generating the image with the 1.5 vae loaded. It improves text generation somewhat.

[ergosumdre]

Thank you! Can I bother you to ask how to do this? There isn't a place in the colab notebook to upload a vae. Sorry, I'm pretty new to this entire space.

[TheLastBen]

this colab has the latest vae included, no need to install it