DreamCache

DreamCache: Finetuning-Free Lightweight Personalized Image Generation via Feature Caching (CVPR'25)

⚙️ Set-up

Create a conda environment dreamcache using

conda env create -f environment.yaml
conda activate dreamcache

⏬ Download

Download the pretained Stable Diffusion 2.1 and move it here: /models/ldm/stable_diffusion_v2

Download the pretrained adapter

Download our Synthetic Dataset. Another good dataset that model pose variation is the Subject200K dataset.

Download DreamBench:

git clone https://github.com/google/dreambooth.git

💻 Training

To run a training job, set main training arguments and run:

bash scripts/train_multi_masked.sh

🚀 Inference & Evals

To generate personalized images modify and run: bash scripts/sample_ref.sh

To sample the entire dreambench dataset first remove background from dreambench and filter according to Kosmos-G. Adjust path in remove_background.py:

bash scripts/remove_dreambench.sh /path/to/dreambooth/dataset
python scripts/remove_background.py 

To generate samples adjust the main arguments and run:

bash scripts/generate_dreambooth_mult.sh

To run evaluations on dreambooth run:

bash scripts/evaluate_dreambooth.sh

Gradio Demo

To run a gradio demo set main parameters in gradio_app.py and then run python gradio_app.py

💐 Acknowledgements

This code repository is based on several prior works: Textual Inversion ViCo Kosmos-G