Command-Line Parameters

This document summarizes the parameters used when running the sldgen.py script. The CLI is organized into sections below, with the most important parameters highlighted first.

Notes:

• Output folders are created automatically under <output-dir>/<target-stem>/<experiment-name>/.
• When --experiment-name is omitted, a timestamp is used.
• When CUDA is unavailable, the code automatically falls back to CPU.
• The target image must exist before running.

Most Important Parameters

These are the parameters you are most likely to tune when running the pipeline.

Parameter	Type	Default	Description
--target	str	required	Path to the image used as the target for single-line drawing generation.
--caption	str	""	Text prompt used for SDS guidance.
--experiment-name	str	generated automatically	Name of the experiment folder. If not provided, a timestamp is used.
--num-iter	int	4000	Number of optimization iterations.
--width	float or str	1.0	Stroke width, or the optim special mode for varying line width.
--fixed-endpoints	flag	False	Keep endpoints fixed during optimization to make it easier to connect drawings.
--n-control-points	int	385	Number of control points used to initialize the line.
--save-interval	int	100	Frequency for saving intermediate results.

Special cases

To generate the particular configurations of single-line drawings presented in the paper, use one of the following commands:

Single-line drawing with varying width

python sldgen.py --target ./data/firefighter.png --width optim

Single-line drawing with fixed endpoints for easier connection

python sldgen.py --target ./data/firefighter.png --fixed-endpoints

All parameters

General Parameters

Parameter	Type	Default	Description
--output-dir	str	./output/	Directory where results are saved.
--experiment-name	str	generated automatically	Name of the experiment folder. If not provided, a timestamp is used.
--use-cpu	flag	False	Force CPU execution even if CUDA is available.
--seed	int	0	Random seed for Python, NumPy, and PyTorch.
--verbose	flag	False	Print loss values during optimization.
--debug	flag	False	Enable debug mode. Print additional debug information.

Target Image Parameters

Parameter	Type	Default	Description
--target	str	required	Target image path. The file must exist.
--object-size-ratio	float	0.75	Maximum size of the object relative to the render size, used to rescale the target.
--render-size	int	512	Output render size in pixels.
--calligraphy	flag	False	Treat the target as a calligraphy image or letter. Changes how the mask is created.

Optimization Parameters

Parameter	Type	Default	Description
--num-iter	int	4000	Total number of optimization iterations.
--lr	float	0.8	Base learning rate for the optimizer.
--save-interval	int	100	Save intermediate outputs every N iterations.

Curve Parameters

Parameter	Type	Default	Description
--sampling-rate	int	5000	Number of points sampled from the curve for rendering.
--no-optimize-cp-weights	flag	optimize_cp_weights = True	Disable optimization of B-spline control-point weights.
--keep-low-weights	flag	prune_low_weights = True	Keep low weights instead of pruning them.
--init-method	str	tsp	Initialization method for the single line: trefoil, contour, or tsp.
--n-control-points	int	385	Number of control points at the start of optimization.
--width	float or str	1.0	Stroke width, or the optim special mode for varying line width.
--fixed-endpoints	flag	False	Keep endpoints fixed during optimization for easier connection between drawings.

SDS Guidance Parameters

Parameter	Type	Default	Description
--caption	str	""	Text prompt used for semantic guidance.
--conditioning-scale	float	0.5	Conditioning scale used in the SDS guidance objective.
--condition	str	depth	Guidance conditioning type, either depth or canny.
--lora-model	str	./SLDgen/guidance/sld-lora.safetensors	Path to the LoRA weights used for guidance.
--lora-weight	float	0.1	Strength of the LoRA adapter. Set to 0 to disable its effect.

Loss Parameters

Parameter	Type	Default	Description
--repulsion-loss-weight	float	0.004	Weight of the repulsion loss.
--sparse-loss-weight	float	2000.0	Weight of the sparsity loss.
--sparse-loss-type	float	1.0	Degree of the sparsity loss; 0.0 corresponds to a standard deviation loss.
--sparse-loss-progressive	str	linear	Progressive sparse-loss mode. Any value other than linear disables the progressive schedule.
--length-shortening-loss-weight	float	0.1	Weight of the length-shortening loss.

Metrics Parameters

Parameter	Type	Default	Description
--aesthetic-predictor-model-path	str	./SLDgen/metrics/aesthetic_predictor_v2_5.pth	Path to the aesthetic predictor model weights.

Fixed Runtime Settings

These values are set internally by SLDgen/config.py after argument parsing and are not exposed as CLI flags.

Setting	Value	Description
diffusion_model	stabilityai/stable-diffusion-3.5-medium	Diffusion model used for guidance.
diffusion_timesteps	1000	Number of diffusion timesteps.
diffusion_guidance_scale	100	Guidance scale used for diffusion sampling.
vae_path	madebyollin/taesd3	VAE used by the guidance pipeline.
multisteps	1	Number of multistep passes.
negative_caption	""	Negative prompt passed to the guidance model.