Controllable Road Marking Generation

Zhiyu; Cai; Yufan Zhang; Ruichen Tan; Zengxiang Lei; Satish Ukkusuri

cs.CVOct 5, 2026

Controllable Road Marking Generation

Authors: Zhiyu, Cai, Yufan Zhang, Ruichen Tan, Zengxiang Lei, Satish Ukkusuri

Organizations: Joey · Purdue University, West Lafayette, IN 47906

Abstract

Lane and road markings provide critical guidance for vehicle navigation and multi-agent coordination, yet authoring them at scale remains a manual workflow that limits quantitative analysis and scenario testing. We introduce Controllable Road Marking Generation, which synthesizes a missing center-region marking layout from a drivable-area mask, optional outer-ring markings, and a textual description. Our benchmark uses deterministic, metadata-derived prompts and three output channels: lane dividers, road dividers, and pedestrian crossings. We develop a conditional bird's-eye-view (BEV) pipeline that combines (i) a text-conditioned latent rectified-flow DiT trained with a topology-aware auxiliary loss, (ii) Gaussian-blurred training targets that stabilize learning of thin, sparse markings, and (iii) Structured Gaussian Render (SGR), a training-free post-process that recovers crisp divider geometry by extracting polylines, fitting cubic Bézier curves, and re-rendering them as anisotropic super-Gaussian primitives. On 4,597 Argoverse~2 test tiles, our system achieves Buffered F1 of 80.8 and clDice of 50.2, compared with 38.8 and 24.6 for an adapted state-of-the-art mask-refinement baseline. On Waymo dataset, it yields 88.0 Buffered F1 and 66.2 clDice. Component ablations show complementary connectivity gains from topology-aware supervision and SGR. Text-editing experiments reveal that stronger guidance improves edit success but also increases changes to non-target structures. We see this framework as a step toward simulation-ready road-marking variation, automated map completion, and early-stage infrastructure design exploration.

Figures & tables

Method	Raster	Generative	Layout	Text	Thin geometry
RoadMark-cGAN [ 7 ]	$\checkmark$	$\checkmark$	$\times$	$\times$	$\triangle$
LSR-DM [ 41 ]	$\checkmark$	$\checkmark$	$\times$	$\times$	$\triangle$
MapDiffusion [ 32 ]	$\times$	$\checkmark$	$\triangle$	$\times$	$\triangle$
LaneDiffusion [ 45 ]	$\times$	$\checkmark$	$\times$	$\times$	$\checkmark$
PolyDiffuse [ 6 ]	$\times$	$\checkmark$	$\triangle$	$\times$	$\triangle$
Ours	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$

Table 1: Capability positioning.

\checkmark

direct support;

\triangle

partial or adjacent capability;

\times

outside scope. Only capabilities are compared in this table.

Figure 1: Overview of the proposed controllable road marking generation pipeline.

Figure 2: Structured Gaussian Render (SGR). Starting from the raw soft raster $\tilde{d}$ , SGR thresholds the divider channels, extracts lane- and road-divider polylines, fits cubic Bézier curves, and re-renders sampled curve primitives with anisotropic super-Gaussian splats. The pedestrian-crossing channel is passed through as a raster mask.

Method	Overall		Lane Div.		Road Div.		Ped. Cross.
Method	F1 $\uparrow$	clDice $\uparrow$	F1 $\uparrow$	clDice $\uparrow$	F1 $\uparrow$	clDice $\uparrow$	F1 $\uparrow$	clDice $\uparrow$
LSR-DM † [ 41 ]	38.8	24.6	37.3	15.6	43.6	27.9	35.4	30.4
Ours w/ CNN deblur	62.7	46.3	50.2	18.5	64.9	37.3	73.0	83.1
Ours w/ SGR deblur	80.8	50.2	74.0	28.4	66.6	38.0	89.7	84.1

Table 2: Structural metrics on

4{,}597

AV2 test tiles, evaluated on the generated center

256{\times}256

region. Per-channel scores use GT-positive tiles; Overall is the macro-average over the three channels. Values are percentages; bold/underline mark best/second-best.

Figure 3: Qualitative text control. Columns show ground truth and edited-prompt generations; green boxes mark the generated center region.

Control family	$N_{\mathrm{elig}}$	Success $\uparrow$	False act. $\downarrow$	fa-MSE $\downarrow$	$\Delta_{\mathrm{tgt}}\uparrow$
Add crosswalk	2,771	49.6	9.2	0.071	$+1{,}772.2$
Delete crosswalk	957	75.1	19.7	0.119	$+4{,}070.4$
Single $\to$ double yellow	1,750	31.0	7.4	0.041	$+29.8$

Table 3: Mask-parsed controllability for single-clause prompt edits on the AV2 test split. Each family is evaluated only on eligible non-trivial tiles (

N_{\mathrm{elig}}

), with

(\mathbf{s},\mathbf{p})

and seed fixed between baseline and edited prompts. Success and False activation are percentages; fa-MSE is a

\tau

-normalized MSE over non-target channel changes, and

\Delta_{\mathrm{tgt}}

is the signed target-channel change. Protocol details are in App. A and App. E .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Semantic case	Example prompt
Two-way crossing tile	“two-way, 2 lanes each direction, solid white dividers, single yellow center line, 4 crosswalks at intersection”
One-way non-crossing tile	“one-way, 1 lane, no double yellow, no crosswalks”
Double-yellow center line	“two-way, 2 lanes each direction, solid white dividers, double yellow center line, 1 crosswalk at intersection”
Lane-type context	“two-way, 2 lanes each direction, solid white dividers, double yellow center line, 2 crosswalks at intersection”

Appendix

Table 4: Representative deterministic prompt strings used for text conditioning.

Dataset	Topology loss	SGR	Buffered F1 $\uparrow$	clDice $\uparrow$
AV2	No	No	80.5	43.4
	No	Yes	81.8	47.2
	Yes	No	78.2	48.1
	Yes	Yes	80.8	50.2
Waymo	No	No	88.6	59.1
	No	Yes	87.3	61.9

Appendix

Table 5: Topology and SGR ablation. Values are percentages; bold marks the best result for each metric within each dataset.

Train data	Steps	Buf. F1 $\uparrow$	clDice $\uparrow$
$0\%$ †	–	3.9	2.4
$10\%$	$18{,}000$	16.2	7.3
$30\%$	$54{,}000$	20.2	9.3
$50\%$ †	$90{,}000$	56.0	39.7
$70\%$ †	$126{,}000$	69.3	47.4
$100\%$	$180{,}000$	80.8	50.2

Appendix

Table 6: Data efficiency under matched-epoch training.

Method	Channel	Buf. F1 $\uparrow$	Buf. IoU $\uparrow$	clDice $\uparrow$
LSR-DM † [ 41 ]	Lane divider	37.3	18.7	15.6
	Road divider	43.6	25.2	27.9
	Pedestrian crossing	35.4	23.4	30.4
Ours w/ CNN deblur	Lane divider	50.2	45.9	18.5
	Road divider	64.9	59.7	37.3
	Pedestrian crossing	73.0	76.5	83.1

Appendix

Table 7: Per-channel structural metrics on GT-positive AV2 test tiles.

Method	Buf. F1 $\uparrow$			clDice $\uparrow$
Method	$r=1$	$r=3$	$r=5$	clDice $\uparrow$
LSR-DM † [ 41 ]	32.8	38.8	41.7	24.6
Ours w/ CNN deblur	59.8	62.7	68.2	46.3
Ours w/ SGR deblur	64.1	73.7	77.5	49.9

Appendix

Table 8: Buffered-F1 sensitivity to tolerance radius.

Asset	Version	License
Argoverse 2 Sensor / Map Dataset	v2.0	CC BY-NC-SA 4.0
CLIP text encoder	ViT-L/14 (OpenAI)	MIT
PyTorch	2.5.1	BSD-3-Clause
diffusers (HuggingFace)	0.30.x	Apache-2.0
xFormers	0.0.28	BSD-3-Clause
NumPy / SciPy / scikit-image	current	BSD-3-Clause

Appendix

Table 9: Existing assets, versions, and licenses.

Figure 4: Additional qualitative comparisons on Argoverse 2 BEV tiles.

Dataset	CFG	Success $\uparrow$	FA $\downarrow$	fa-MSE $\downarrow$	$\Delta_{\mathrm{tgt}}\uparrow$
AV2	1	7.7	6.0	0.0170	$+26.8$
	3	24.3	12.7	0.0406	$+66.1$
	5	29.3	17.3	0.0707	$+74.1$
	7	35.3	20.7	0.0812	$+98.5$
Waymo	1	12.1	4.4	0.0110	$+45.9$
	3	28.7	7.4	0.0369	$+71.9$

Appendix

Table 10: Fixed-seed classifier-free guidance sweep on 300 eligible tiles per dataset, evaluated on raw outputs before SGR. Success and false activation (FA) are percentages.

\Delta_{\mathrm{tgt}}

denotes direction-normalized target change.

Dataset	Overall		Lane Div.		Road Div.		Ped. Cross.
Dataset	F1 $\uparrow$	clDice $\uparrow$	F1 $\uparrow$	clDice $\uparrow$	F1 $\uparrow$	clDice $\uparrow$	F1 $\uparrow$	clDice $\uparrow$
nuScenes-SG †	30.4	8.1	40.7	4.6	33.1	8.7	17.4	11.0
AV2	80.8	50.2	74.0	28.4	66.6	38.0	89.7	84.1
Waymo	88.0	66.2	89.5	55.9	85.7	56.5	88.9	86.3

Appendix

Table 11: Cross-dataset structural metrics for Ours w/ SGR deblur, evaluated on the generated center

256{\times}256

region. Results are reported on

261

nuScenes-SG held-out tiles,

4{,}597

AV2 test tiles, and

1{,}000

Waymo test tiles. Per-channel scores use GT-positive tiles; Overall is the macro-average over the three channels. Values are percentages.

Explore similar work

ControlMap: Controllable High-Definition Map Generation for Traffic Scenario Simulation
Jun 14, 2026Marwan Farag, Steffen Wäldele, Yu YaoLane-Level Map Generation Latent Diffusion Model
University of Stuttgart, Stuttgart, Germany · Robert Bosch GmbH, Renningen, Germany · Motional, Inc
MapAgent: An Industrial-Grade Agentic Framework for City-scale Lane-level Map Generation
Jun 3, 2026Deguo Xia, Zihan Li, Haochen Zhao +6Lane-Level Map Generation Autonomous Driving
Tsinghua University Baidu Beijing, China · University of Macau Macao, China · Institute of Information Engineering, Chinese Academy of Sciences Beijing, China +2
MapDreamer: Aerial Imagery Conditioned Latent Diffusion for Lane-Level Map Generation
Jul 1, 2026Julian Brandes, Philipp Crocoll, Wolfram BurgardLane-Level Map Generation Autonomous Driving
Department CSAI at University of Technology Nuremberg, Nuremberg, Germany · Robert Bosch GmbH, Stuttgart

Jun 14, 2026, cs.ROJ/K move · Enter open · S save

ControlMap: Controllable High-Definition Map Generation for Traffic Scenario Simulation

Marwan Farag, Steffen Wäldele, Yu Yao

University of Stuttgart, Stuttgart, Germany · Robert Bosch GmbH, Renningen, Germany · Motional, Inc

Lane-Level Map Generation Latent Diffusion Model

Open paper AskPDF arXiv

Simulation is central to validating autonomous driving systems, yet current pipelines are limited by insufficient scenario diversity due to costly High Definition (HD) map creation. Scaling HD maps requires expensive data collection and manual processing. Moreover, existing generative models lack the fine-grained control necessary to target specific road topologies during generation. This paper presents a data-driven pipeline for controllable HD map generation using latent diffusion and ControlNet for spatial conditioning. To our knowledge, we are the first to inject spatial guidance signals into a diffusion model for HD map synthesis. Furthermore, our model supports adjustable conditioning strength through classifier-free guidance and city-level style transfer via city label conditioning. To complement existing metrics, we introduce two novel metrics to evaluate adherence to the control signal and similarity to ground-truth maps. Experiments demonstrate that our model generates realistic HD maps that faithfully follow input road topologies while accurately preserving city-specific details.

Abstract

Explore similar work

ControlMap: Controllable High-Definition Map Generation for Traffic Scenario Simulation

MapAgent: An Industrial-Grade Agentic Framework for City-scale Lane-level Map Generation

MapDreamer: Aerial Imagery Conditioned Latent Diffusion for Lane-Level Map Generation