ProFusion (with an encoder pre-trained on a large dataset such as CC3M) can be used to efficiently construct customization dataset, which can be used to train a tuning-free customization assistant (CAFE).
Given a testing image, the assistant can perform customized generation in a tuning-free manner. It can take complex user-input, generate text explanation and elaboration along with image, without any fine-tuning.
Results from CAFE
Results from CAFE
Results from ProFusion
ProFusion is a framework for customizing pre-trained large-scale text-to-image generation models, which is Stable Diffusion 2 in our examples.
Illustration of the proposed ProFusion
With ProFusion, you can generate infinite number of creative images for a novel/unique concept, with single testing image, on single GPU (~20GB are needed when fine-tune with batch size 1).
Results from ProFusion
-
Install dependencies (we revised original diffusers);
cd ./diffusers pip install -e . cd .. pip install accelerate==0.16.0 torchvision transformers==4.25.1 datasets ftfy tensorboard Jinja2 regex tqdm joblib -
Initialize Accelerate;
accelerate config -
Download a model pre-trained on FFHQ;
-
Customize model with a testing image, example is shown in the notebook test.ipynb;
If you want to train a PromptNet encoder for other domains, or on your own dataset.
-
First, prepare an image-only dataset;
- In our experiments on human face domain, we use FFHQ. Our pre-processed FFHQ can be found at google drive link;
- We also trained an encoder on CC3M, which leads to good customization results on arbitrary downstream domain;
-
Then, run
accelerate launch --mixed_precision="fp16" train.py\ --pretrained_model_name_or_path="stabilityai/stable-diffusion-2-base" \ --train_data_dir=./images_512 \ --max_train_steps=80000 \ --learning_rate=2e-05 \ --output_dir="./promptnet" \ --train_batch_size=8 \ --promptnet_l2_reg=0.000 \ --gradient_checkpointing
@article{zhou2023enhancing,
title={Enhancing Detail Preservation for Customized Text-to-Image Generation: A Regularization-Free Approach},
author={Zhou, Yufan and Zhang, Ruiyi and Sun, Tong and Xu, Jinhui},
journal={arXiv preprint arXiv:2305.13579},
year={2023}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent