DiffSketcher

This is a Hugging Face implementation of DiffSketcher: Text Guided Vector Sketch Synthesis through Latent Diffusion Models.

Model Description

DiffSketcher is a novel approach for synthesizing vector sketches from text prompts by leveraging the power of latent diffusion models. It extracts cross-attention maps from a pre-trained text-to-image diffusion model and uses them to guide the optimization of vector sketches.

Usage

from diffusers import DiffusionPipeline

# Load the pipeline
pipeline = DiffusionPipeline.from_pretrained("jree423/diffsketcher")

# Generate a vector sketch
result = pipeline(
    prompt="A beautiful sunset over the mountains",
    negative_prompt="ugly, blurry",
    num_paths=96,
    token_ind=4,
    num_iter=800,
    guidance_scale=7.5,
    width=1.5,
    seed=42
)

# Access the SVG string and rendered image
svg_string = result["svg"]
image = result["image"]

# Save the SVG
with open("sunset_sketch.svg", "w") as f:
    f.write(svg_string)

# Save the image
image.save("sunset_sketch.png")

Parameters

prompt (str): The text prompt to guide the sketch generation.
negative_prompt (str, optional): Negative text prompt for guidance.
num_paths (int, optional): Number of paths to use in the sketch. Default is 96.
token_ind (int, optional): Token index for attention. Default is 4.
num_iter (int, optional): Number of optimization iterations. Default is 800.
guidance_scale (float, optional): Scale for classifier-free guidance. Default is 7.5.
width (float, optional): Stroke width. Default is 1.5.
seed (int, optional): Random seed for reproducibility.
return_dict (bool, optional): Whether to return a dict or tuple. Default is True.
output_type (str, optional): Output type, one of "pil", "np", or "svg". Default is "pil".

Citation

@article{xing2023diffsketcher,
  title={DiffSketcher: Text Guided Vector Sketch Synthesis through Latent Diffusion Models},
  author={Xing, Ximing and Xie, Chuang and Qiao, Yu and Xu, Hongteng},
  journal={arXiv preprint arXiv:2306.14685},
  year={2023}
}