ConsistencyDifFeature

Setup

conda create -n cdhf python=3.10
conda activate cdhf
pip install -r requirements.txt

Folder Structure

configs

e.g. train.yaml

parameters:

save_timestep: choose which timestep to aggregate
lcm_model_name: choose the diffusion model
extract_mode:"iid_noise_denoise" means add one step noise

train_aggregation_network.py

set config path as train.yaml, then run train_aggregation_network.py

test_aggregation_network.py

set config path as test.yaml, then run test_aggregation_network.py

get the pca outcome

set config path as pca.yaml run pca_visualization.py in utils

try the correspondence task in any pair of image

using notebook run demo.ipynb