LayerTracer: Cognitive-Aligned Layered SVG Synthesis via Diffusion Transformer
Yiren Song, Danze Chen, and Mike Zheng Shou
Show Lab, National University of Singapore

git clone https://github.com/showlab/LayerTracer.git
cd LayerTracer

conda create -n layertracer python=3.11.10
conda activate layertracer

pip install --upgrade -r requirements.txt

You can download the trained checkpoints of LoRA for text2sequence inference. Below are the details of available models:

According to the actual situation, replace all model paths, file paths, and parameters in scripts/text2sequence.sh.

chmod +x scripts/text2sequence.sh
scripts/text2sequence.sh

NOTICE

When generating 9-grid images using prompts, please modify your prompt according to the examples provided below to avoid generating irrelevant content. Add this prefix to your prompt:

# nine-grid icon
sks, This is a nine square grid image that describes the process of creating SVG icon images, ...
# or
sks, a set of nine icons, each representing ...

# four-grid emoji
sks, This is a four square grid image that describes the process of creating SVG emoji images, ...

Use our scripts/lora_merge.sh template script to merge the LoRA （flux_lora_icon_blackline and flux_lora_emoji) to flux.1 checkpoints for further recraft training. Note that the merged model may take up around 50GB of your memory space.

chmod +x scripts/lora_merge.sh
scripts/lora_merge.sh

According to the actual situation, replace all model paths, file paths, and parameters in scripts/image2sequence.sh.

The icon_lora_weights is a Recraft model used for generating 9-grid black line icon images. Simply replace the corresponding paths in scripts/image2sequence.sh to use it.

chmod +x scripts/image2sequence.sh
scripts/image2sequence.sh

python layertracer_icon_9grid.py --input input/icon.png --output output/

• --input : Input image path.
• --output : Output directory.
• --colormode: choices=['color', 'binary'], Color mode for SVG conversion.
• --mode : choices=['spline', 'polygon', 'none'], Tracing mode for SVG conversion.
• --filter_speckle : Speckle filter threshold.
• --color_precision : Color precision for SVG conversion.
• --corner_threshold: Corner detection threshold.
• --length_threshold : Length threshold for path simplification, in [3.5, 10].
• --splice_threshold : Splice threshold for path merging.
• --path_precision : Path precision for SVG conversion.

You can customize the parameters, just like the above process:

python layertracer_emoji_4grid.py --input input/emoji.png --output output/

• 4grid_emoji_icon
• 9grid_icon_blackline

@inproceedings{Song2025LayerTracerCL,
  title={LayerTracer: Cognitive-Aligned Layered SVG Synthesis via Diffusion Transformer},
  author={Yiren Song and Danze Chen and Mike Zheng Shou},
  year={2025},
  url={https://api.semanticscholar.org/CorpusID:276094351}
}