How did you train it so that it generates crayon drawing only without having to mention any prefix like "crayon drawing of" to the prompt?

#1
by kteedle - opened

I ran a notebook by TheLastBen i found online for kohya dreambooth which didnt have any option to put instance prompt i think is that the problem? It said in the image uploading section of the notebook that - "If you have 10 pictures of yourself, simply select them all and rename only one to the chosen identifier for example : phtmejhn, the files would be : phtmejhn (1).jpg, phtmejhn (2).png ....etc then upload them, do the same for other people or objects with a different identifier, and that's it."
But i was not training for a person i wanted to train for a style and i didn't name the files anything i just gave them blip captions.
But when i trained and tested it needed a prompt separately explaining the desired style type i want otherwise was generating normal realistic images, which defeats the whole purpose of training it on my style of images.

So please let me know how to train dreambooth or lora to get any desired style images like you did without having to write any prefix in the prompt. The generations look so much like your crayon drawings! Thank you in advance πŸ™ @ostris

this is the TheLastBen notebook i used for training - https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb

kteedle changed discussion status to closed
kteedle changed discussion status to open

Hey sorry forgot to ask if you put regularization images too? if yes then how many regularization images did you put of your crayon drawings? Kindly let me know Thankyou

Sign up or log in to comment