README.md

Scale-Adaptive Diffusion Model for Complex Sketch Synthesis

Jijin Hu · Ke Li · Yonggang Qi · Yi-Zhe Song
Beijing University of Posts and Telecommunications   |   SketchX, CVSSP, University of Surrey

[OpenReview]

News

• [2024.06.16] Release training and inference code, checkpoints.

Installaiton

conda env create -f environment.yml
conda activate ada_guidance

Download Checkpoints

Here are the download links for each model checkpoint:

• 64x64 unconditional diffusion: 64x64_uncond_diffusion.pt
• 64x64 classifier: 64x64_classifier.pt

Inference

For sampling from the 64x64 classifier-guided diffusion model with adaptive guidance scale, 250 step DDIM:

sh classifier_sample.sh

Training Models

train unconditional diffusion models:

sh diffusion_train.sh

train a noised classifier:

sh classifier_train.sh

Results

Quantitative comparison results on QuickDraw Dataset. The best and second best are color-coded in red and blue, respectively.

Model	Random 30 Categories					345 Categories
	FID↓	CLIP-Score↑	CLIP-Fine(%)↑	Prec↑	Rec↑	FID↓	Prec↑	Rec↑
SketchRNN	8.15	0.59	52.67	0.37	0.22	10.32	0.26	0.24
SketchHealer	5.85	0.63	51.51	0.67	0.12	--	--	--
SketchAA	5.98	0.59	50.41	0.51	0.17	--	--	--
SketchKnitter	7.05	0.55	43.15	0.41	0.19	--	--	--
ChiroDiff	4.75	0.59	53.16	0.64	0.18	3.17	0.58	0.25
StyleGAN2	4.12	0.67	53.39	0.55	0.24	2.93	0.63	0.27
DDIM	4.08	0.67	54.19	0.71	0.30	2.85	0.74	0.31
CFDG	3.75	0.68	54.86	0.68	0.32	2.64	0.73	0.36
Ours	3.08	0.68	55.54	0.68	0.35	2.51	0.72	0.39

Figure (b) shows some of our results. For more detailed results, please see our paper and supplementary materials.

Acknowledgements

This project is developped on the codebase of guided-diffusion. We appreciate this great work!

Citation

If you find this codebase useful for your research, please use the following entry.

@inproceedings{huscale,
  title={Scale-Adaptive Diffusion Model for Complex Sketch Synthesis},
  author={Hu, Jijin and Li, Ke and Qi, Yonggang and Song, Yi-Zhe},
  booktitle={The Twelfth International Conference on Learning Representations}
}