PyTorch Code for "What's in a Latent? Leveraging Diffusion Latent Space for Domain Generalization" (ICCV 2025)
Xavier Thomas, Deepti Ghadiyaram
- GUIDE is implemented as an algorithm in the
domainbed/algorithms.pyfile. - See Precompute Features for instructions on how to save the features (
${\Psi}$ ) to disk.
- Run
run_precompute.shby setting the layer and model arguments. - The features are saved in the
domainbed/saved_featsdirectory.
- For GUIDE-SD-2.1 on TerraIncognita dataset, run:
- Get Stable Diffusion 2.1 Features at up_ft:1 and timestep 50 for TerraIncognita
bash run_precompute.sh- Run DomainBed on a single environment for the above setting
python3 -m domainbed.scripts.train_precompute\
--data_dir=domainbed/data/\
--algorithm GUIDE\
--dataset TerraIncognita\
--test_env 3\
--hparams '{"model_name": "stabilityai/stable-diffusion-2-1-base", "feature_model": "diffusion", "timestep": 50, "num_clusters": 5}'- Create a sweep
python -m domainbed.scripts.sweep launch \
--data_dir=domainbed/data \
--output_dir=GUIDE_SD_TI \
--command_launcher multi_gpu \
--algorithms GUIDE \
--datasets TerraIncognita \
--n_hparams 1 \
--hparams '{"model_name": "stabilityai/stable-diffusion-2-1-base", "feature_model": "diffusion", "timestep": 50, "num_clusters": 5}'\
--n_trials 1 \
--steps 5001 \
--skip_confirmation \This code is built on top of Domainbed, visit Domainbed for more details on running training sweeps, hyperparameter configurations, etc.
@misc{thomas2025whatslatentleveragingdiffusion,
title={What's in a Latent? Leveraging Diffusion Latent Space for Domain Generalization},
author={Xavier Thomas and Deepti Ghadiyaram},
year={2025},
eprint={2503.06698},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2503.06698},
}