Accelerating Stable Diffusion Inference via Selective Caching

Installation

pip install -e .

How to run the experiments

For individual generation:

bash run.sh # using basic deepcache
export ENABLE_SMART_INTERVAL=1; bash run.sh  # enable smart-internval opt
export ENABLE_MASK=1; bash run.sh # enable mask opt

Run benchmarking

pip install diffusers transformers open_clip_torch
python generate.py --dataset coco2017 --layer 0 --block 0 --update_interval 10 --uniform --steps 50 --batch_size 16 
export ENABLE_SMART_INTERVAL=1; python generate.py --dataset coco2017  --original --steps 50 --batch_size 16 # For original pipeline

python score_is.py PATH_TO_SAVED_IMAGES

Ackownledgement

• Based on DeepCache