Skip to content

Latest commit

 

History

History
98 lines (69 loc) · 3.83 KB

README.md

File metadata and controls

98 lines (69 loc) · 3.83 KB
Raw

Diffusion Content Replication study

This repo contains code for two papers.

  1. Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models (CVPR'23) - paper link
  2. Understanding and Mitigating Copying in Diffusion Models (NeurIPS'23) - paper link

alt text

Set up

Install conda environment

conda env create -f env.yaml
conda activate diffrep

We used RTX-A6000 machines to train the models. For inference or to compute the metrics, smaller machines will do.

Finetuning a model

accelerate launch diff_train.py \
  --pretrained_model_name_or_path stabilityai/stable-diffusion-2-1 \
  --instance_data_dir <training_data_path> \
  --resolution=256 --gradient_accumulation_steps=1 --center_crop --random_flip \
  --learning_rate=5e-6 --lr_scheduler constant_with_warmup \
  --lr_warmup_steps=5000  --max_train_steps=100000 \
  --train_batch_size=16 --save_steps=10000 --modelsavesteps 20000 --duplication <duplication_style>  \
  --output_dir=<path_to_save_model> --class_prompt <conditioning_style> --instance_prompt_loc <path_to_captions_json>
  • <duplication_style> options are nodup,dup_both,dup_image.
  • <conditioning_style> options are nolevel,classlevel,instancelevel_blip,instancelevel_random.
  • To train a model with mitigation, set --trainspecial <traintime_mitigation_strategy>. The available options are allcaps,randrepl,randwordadd,wordrepeat.
  • For gaussian noise addition in training, set --rand_noise_lam to a non-zero value in range [0,1].

Inference from a finetuned model

python diff_inference.py --modelpath <path_to_finetuned_model> -nb <number_of_inference_generations>

Computing metrics

This script computes similairity scores, fid scores and a few other metrics. Logged to wandb.

python diff_retrieval.py --arch resnet50_disc --similarity_metric dotproduct --pt_style sscd --dist-url 'tcp://localhost:10001' --world-size 1 --rank 0 --query_dir <path_to_generated_data> --val_dir <path_to_training_data>

Mitigation strategy

For train time strategies, train with --trainspecial option and infer and compute metrics as shown below

For inference time strategies,

python sd_mitigation.py --rand_noise_lam 0.1 --seed 2
python sd_mitigation.py --rand_augs rand_word_repeat --seed 2
python sd_mitigation.py --rand_augs rand_word_add --seed 2
python sd_mitigation.py --rand_augs rand_numb_add --seed 2

Data

Download the LAION-10k split here.

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.

Cite us

Citation for Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models

@inproceedings{somepalli2023diffusion,
  title={Diffusion art or digital forgery? investigating data replication in diffusion models},
  author={Somepalli, Gowthami and Singla, Vasu and Goldblum, Micah and Geiping, Jonas and Goldstein, Tom},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6048--6058},
  year={2023}
}

Citation for Understanding and Mitigating Copying in Diffusion Models

@article{somepalli2023understanding,
  title={Understanding and Mitigating Copying in Diffusion Models},
  author={Somepalli, Gowthami and Singla, Vasu and Goldblum, Micah and Geiping, Jonas and Goldstein, Tom},
  journal={arXiv preprint arXiv:2305.20086},
  year={2023}
}