how many prompts we need for training this model

#1
by mayukitan - opened

Hi, first thank you for providing excellent work to create a one-step image generation model. We want to train a customized model based on the architecture of Swiftbrush. Having reviewed your GitHub page, we have several questions regarding the training phase

  1. What are the formats of the prompts.txt file we need to create since we didn't find a sample file in your git repo
  2. How many prompts do we need for training a model that can generate similar quality and FID score image like the pre-trained one you have
  3. Can we replace the base model for training, right now you put stable-diffusion-2-1-base as the base mode. Should we keep using this one or can we change to other diffusion models?

Looking forward to your response, thanks

Hi, sorry for the late reply. Usually I dont check this site much. I will answer each of your questions as below:

  1. Just a .txt file containing a list of prompts like following:

    a dog
    a man holding an apple
    a painting of a cat

  2. We deduplicate the prompts from JourneyDB and train the model with that. It’s about 1.3M prompts and takes 4 days to train using an A100 gpu

  3. Yes you can use any other multi-step diffusion models as you want to improve the quality or have any effect that you desire. But you may need some hyper-parameter tuning.

Sign up or log in to comment