Detecting, Explaining, and Mitigating Memorization in Diffusion Models

Official repo for Detecting, Explaining, and Mitigating Memorization in Diffusion Models.

If you have any questions, feel free to email Yuxin (ywen@umd.edu).

Dependencies

• PyTorch == 1.13.0
• transformers == 4.30.2
• diffusers == 0.18.2
• accelerate == 0.21.0
• datasets

Play around our methods

We provide detecting, explaining, and (inference) mitigating methods separately in the Jupyter notebooks within the examples/ folder. You can directly interact with our methods without the need to download any datasets or fine-tune any models.

To reproduce the results in our paper

Detect memorization

1. Run detection:

memorized prompts

python detect_mem.py --run_name memorized_prompts --dataset examples/sdv1_500_memorized.jsonl --end 500 --gen_seed 0

non-memorized prompts

python detect_mem.py --run_name non_memorized_prompts --dataset Gustavosta/Stable-Diffusion-Prompts --end 500 --gen_seed 0

2. Visualize resuts: you could use examples/det_mem_viz.ipynb.

Explain memorization

You could check out examples/token_wise_significance.ipynb.

Mitigate memorization

For our mitigation experiments, we fine-tune stable diffusion models on LAION to experiment with memorized examples with more meaningful prompts.

You could either use our fine-tuned models by downloading the pre-fine-tuned checkpoint and memorized images or fine-tune the model with the following instruction:

Fine-tune Stable Diffusion model

accelerate launch --mixed_precision=fp16 train_text_to_image.py --dataset=$MEM_DATA --non_mem_dataset=$NON_MEM_DATA --output_dir=finetuned_checkpoints/v1-20000 --pretrained_model_name_or_path=CompVis/stable-diffusion-v1-4 --end=200 --repeats=200 --non_mem_ratio=3 --use_ema --resolution=512 --center_crop --train_batch_size=8 --gradient_accumulation_steps=1 --gradient_checkpointing --max_train_steps=20000 --checkpointing_steps=20000 --learning_rate=1e-05 --max_grad_norm=1 --lr_scheduler=constant --lr_warmup_steps=0

dataset and non_mem_dataset should provide the path to the dataset you want to memorize and the base dataset respectively.

By default, the code assumes the data folder that contains the image and text pairs looks like:

-- daataset
    |-- 00000001.jpg
    |-- 00000001.txt
    |-- 00000002.jpg
    |-- 00000002.txt
    |-- ...

Inference-time mititgation

You may want to download the SSCD checkpoint first here.

No mitigation:

python inference_mem.py --run_name no_mitigation --unet_id finetuned_checkpoints/checkpoint-20000/unet --dataset memorized_images --end 200 --gen_seed 0 --reference_model ViT-g-14 --with_tracking

Random token addition baseline:

python inference_mem.py --run_name add_rand_word --unet_id finetuned_checkpoints/checkpoint-20000/unet --dataset memorized_images --end 200 --gen_seed 0 --prompt_aug_style rand_word_add --repeat_num 4 --reference_model ViT-g-14 --with_tracking

Ours:

python inference_mem.py --run_name ours --unet_id finetuned_checkpoints/checkpoint-20000/unet --dataset memorized_images --end 200 --gen_seed 0 --optim_target_loss 3 --optim_target_steps 0 --reference_model ViT-g-14 --with_tracking

Training-time mitigation

No mitigation:

accelerate launch --mixed_precision=fp16 train_text_to_image.py --dataset=$MEM_DATA --non_mem_dataset=$NON_MEM_DATA --output_dir=finetuned_checkpoints/no_mitigation --pretrained_model_name_or_path=CompVis/stable-diffusion-v1-4 --end=200 --repeats=200 --non_mem_ratio=3 --use_ema --resolution=512 --center_crop --train_batch_size=8 --gradient_accumulation_steps=1 --gradient_checkpointing --max_train_steps=20000 --checkpointing_steps=20000 --learning_rate=1e-05 --max_grad_norm=1 --lr_scheduler=constant --lr_warmup_steps=0

Random token addition baseline:

accelerate launch --mixed_precision=fp16 train_text_to_image.py --dataset=$MEM_DATA --non_mem_dataset=$NON_MEM_DATA --output_dir=finetuned_checkpoints/add_rand_word --pretrained_model_name_or_path=CompVis/stable-diffusion-v1-4 --end=200 --repeats=200 --non_mem_ratio=3 --use_ema --resolution=512 --center_crop --train_batch_size=8 --gradient_accumulation_steps=1 --gradient_checkpointing --max_train_steps=20000 --checkpointing_steps=20000 --learning_rate=1e-05 --max_grad_norm=1 --lr_scheduler=constant --lr_warmup_steps=0 --prompt_aug_style rand_word_add --repeat_num 1

Ours:

accelerate launch --mixed_precision=fp16 train_text_to_image.py --dataset=$MEM_DATA --non_mem_dataset=$NON_MEM_DATA --output_dir=finetuned_checkpoints/ours --pretrained_model_name_or_path=CompVis/stable-diffusion-v1-4 --end=200 --repeats=200 --non_mem_ratio=3 --use_ema --resolution=512 --center_crop --train_batch_size=8 --gradient_accumulation_steps=1 --gradient_checkpointing --max_train_steps=20000 --checkpointing_steps=20000 --learning_rate=1e-05 --max_grad_norm=1 --lr_scheduler=constant --lr_warmup_steps=0 --hard_threshold 2

Then, you can use the fnference-time mititgation script to evaluate the memorization:

python run_mem.py --run_name baseline --unet_id finetuned_checkpoints/checkpoint-20000/unet --dataset memorized_images --end 200 --gen_seed 0 --reference_model ViT-g-14 --with_tracking

Calculate memorization with sdv1_500_memorized.jsonl

You can first download all the groundtruth images here.

Then, run

python inference_mem.py --run_name no_mitigation --dataset sdv1_500_mem_groundtruth --end 500 --gen_seed 0 --reference_model ViT-g-14 --with_tracking

Suggestions and pull requests are welcome!

Cite our work

If you find this work useful, please cite our paper:

@inproceedings{
wen2024detecting,
title={Detecting, Explaining, and Mitigating Memorization in Diffusion Models},
author={Yuxin Wen and Yuchen Liu and Chen Chen and Lingjuan Lyu},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=84n3UwkH7b}
}