DiffSketch: Plug-and-Play Diffusion Features for Zero-Shot Line-Drawing-to-Sketch Synthesis
Stable Diffusion Implementation, our method is built on PnP-Diffusion
To generate Facial Sketch given line drawings, please follow these steps:
Setup
Our codebase is built on CompVis/stable-diffusion and has shared dependencies and model architecture.
Creating a Conda Environment
conda env create -f environment.yaml
conda activate pnp-diffusion
Downloading StableDiffusion Weights
Download the StableDiffusion weights from the CompVis organization at Hugging Face
(download the sd-v1-4.ckpt file), and link them:
mkdir -p models/pnp_ldm/stable-diffusion-v1/
ln -s <path/to/model.ckpt> models/pnp_ldm/stable-diffusion-v1/model.ckpt
Setting Experiment Root Path
The data of all the experiments is stored in a root directory.
The path of this directory is specified in configs/pnp/setup.yaml, under the config.exp_path_root key.
Hardware
The code was tested with a single GPU (NVIDIA GeForce RTX 3090) with 24GB of memory.
Running Synthesis
Note that the data for testing is already provided in the repository (in data directory).
Running on a Single Image
To run the model on a single image, and test with hyperparameters, run the notebook pnp_sketch.ipynb.
Some important parameters to set: self_attn_output_block_indices, out_layers_output_block_indices, to specify the layers to use for the self-attention and the output layers, respectively; prompts, used to specify the prompts to use for the diffusion process; img_path, the path to the input image; exp_config.scale, used to specify the scale of unconditional guidance (a small value encourage the model to generate a sketch that is close to the input image). More parameters can be found in configs/pnp_refine_all.yaml.
Running on a Dataset
To run the model on a dataset, run the python script pnp_sketch.py.
python pnp_sketch.py
All required hyperparameters are specified in configs/pnp_refine_all.yaml.
Gather Results
To gather the results, run the notebook gather.ipynb.
This notebook will create a directory result in the root directory, and will gather the results of the experiments in this directory for submission.
Additional Results
We provide the comparison with DiffStyle and the ablation study as below.
Citation
@article{pnpDiffusion2022,
title={Plug-and-Play Diffusion Features for Text-Driven Image-to-Image Translation},
author={Tumanyan, Narek and Geyer, Michal and Bagon, Shai and Dekel, Tali},
journal={arXiv preprint arXiv:2211.12572},
year={2022}
}