Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
|
|
1 year ago | |
|---|---|---|
| dreambooth | 1 year ago | |
| LICENSE | 1 year ago | |
| README.md | 1 year ago | |
本项目参考自 JDiffusion 的 DreamBooth-Lora 。
按照 JDiffusion 的环境安装一节 安装必要的依赖。
接着设置运行脚本的权限:
chmod u+x ./dreambooth/*.sh
A 下载到 ./A 下train_all.sh 中的 HF_HOME 设置为本地模型路径, root 设置为项目所在目录, BASE_INSTANCE_DIR 设置为数据集对应的目录,GPU_COUNT 设置为对应可用的显卡数量,MAX_NUM 设置为数据集中的风格个数;cd ./dreambooth/, 运行 bash train_all.sh 即可训练。保存的模型会存放至 ./dreambooth/results/prompt_v1_color_test1/style_[训练epoch数]epoch 目录下,例如:./dreambooth/results/prompt_v1_color_test1/style_300epoch。test_all.sh 中的 HF_HOME 设置为本地模型路径,将 run_all.py 中的 root 设置为项目所在目录, dataset_root 修改为数据集对应的目录,将 max_num 修改为数据集中的风格个数;cd ./dreambooth/,运行 bash test_all.sh 进行推理。模型生成的图片会输出到 ./dreambooth/results/prompt_v1_color_test1/outputs_[保存点训练epoch数]ckpt_[推理轮数]steps_[种子值]seed 文件夹下,例如:./dreambooth/results/prompt_v1_color_test1/outputs_300ckpt_500steps_76587seed。@inproceedings{ruiz2023dreambooth,
title={Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation},
author={Ruiz, Nataniel and Li, Yuanzhen and Jampani, Varun and Pritch, Yael and Rubinstein, Michael and Aberman, Kfir},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2023}
}
Jittor 2024 Competition Track 2, Rank 24 in A List
Python Shell