deepinfra/dreambooth
A new approach for personalizing text-to-image diffusion models to user needs has been proposed, enabling the synthesis of novel images of a specific subject in different contexts. This method involves fine-tuning a pretrained model with a few input images of the subject and using a unique identifier to embed the subject in the output domain. As a result, fully-novel photorealistic images of the subject can be synthesized in diverse scenes, poses, views, and lighting conditions that are not present in the reference images.
A new approach for personalizing text-to-image diffusion models to user needs has been proposed, enabling the synthesis of novel images of a specific subject in different contexts. This method involves fine-tuning a pretrained model with a few input images of the subject and using a unique identifier to embed the subject in the output domain. As a result, fully-novel photorealistic images of the subject can be synthesized in diverse scenes, poses, views, and lighting conditions that are not present in the reference images.
Input
The prompt used to describe your training images. Should be something like:'a photo of [identifier] [class noun]'. where [identifier] should be something unique
The prompt used to describe your class. Should be something like:'a photo of a [class noun]'. where [class_noun] describe a class of instance
Please upload a zip file
The name of the model to be saved. Must start with your github username
The maximum number of training steps to run (Default: 1000)
The batch size to use during training (Default: 1)
Whether to center crop the images before resizing to resolution 2
Initial learning rate (after the potential warmup period) to use. (Default: 0.000001)
The scheduler type to use during training. 6
Number of steps for the warmup in the lr scheduler (Default: 0)
Whether or not to use 8-bit Adam from bitsandbytes. 2
The beta1 parameter for the Adam optimizer. (Default: 0.9)
The beta2 parameter for the Adam optimizer. (Default: 0.999)
Weight decay to use (Default: 0.01)
Epsilon value for the Adam optimizer (Default: 1e-8)
Max gradient norm. (Default: 1)
Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass. 2
You need to login to use this model
Output
Model Name:
my_user_name/my_cool_model
Version:
90lkjfslkf5lkdjflskjfsdkl
