UniMapGen: Unified Map Generation

• UniMapGen is a generative framework for large-scale, lane-level map construction that treats map generation as conditional sequence prediction using BEV, PV, text prompts, and previous-map states.
• It converts vector maps into discrete token sequences with dynamic ordering and state updates, ensuring global continuity and reducing the need for extensive post-processing.
• Empirical results on OpenSatMap20 show state-of-the-art performance, validating its multi-modal conditioning and innovative approach to continuous map generation.

Searching arXiv for the core UniMapGen paper and closely related unified map-generation/encoding work.

UniMapGen is a generative framework for large-scale, lane-level map construction from multi-modal data. In its canonical formulation, it treats map construction as an autoregressive sequence generation problem over serialized vector maps, conditioned on bird’s-eye-view satellite imagery, perspective-view ground images, text prompts, and a previous-state map. The framework is designed to address two limitations identified in satellite-based map construction: inherent satellite drawbacks such as occlusions and outdatedness, and inefficient vectorization from perception-based methods that yields discontinuous and rough roads requiring extensive post-processing. Its reported contributions are threefold: representing lane lines as a discrete sequence, supporting multi-modal inputs with dynamic selection among BEV, PV, and text prompt, and introducing a state update strategy for global continuity and consistency. On OpenSatMap20 validation, UniMapGen reports state-of-the-art results, including $\text{AP}^{C}_{0.9}=29.17$, $\text{AP}^{M}=8.38$, $\text{AP}^{M}_{50}=21.67$, and mIoU $=41.81$ (Yuan et al., 26 Sep 2025).

1. Problem formulation and design objective

Large-scale, lane-level map construction is presented as a foundational requirement for autonomous driving and advanced navigation because vehicles require precise lane geometries and connectivity for localization, trajectory planning, and lane-level routing (Yuan et al., 26 Sep 2025). The starting point is the cost structure of traditional HD mapping: specialized survey vehicles, complex offline pipelines, and significant manual annotation and quality assurance. Satellite-based alternatives improve coverage and efficiency, but the paper isolates two technical bottlenecks.

The first bottleneck is the inherent weakness of satellite imagery. Roads may be occluded by buildings, trees, or bridges, and the imagery can be outdated relative to current road layouts. The second bottleneck is the vectorization strategy of perception-based methods. Segmentation-based pipelines operate patch-wise and then rely on skeletonization, thinning, and grouping, often producing discontinuous roads at patch boundaries and rough geometry. Fixed-point vector methods, such as approaches that use the same number of sampled points for all polylines, under-sample long curves and over-sample short segments, again without guaranteeing global continuity (Yuan et al., 26 Sep 2025).

UniMapGen redefines the task as conditional sequence generation. For patch or state $n$, the framework predicts a set of lane vectors and attributes:

$$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$

with each vector

$\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$

The iterative mapping problem is then written as

$$G_n = \text{UniMapGen}(I^{BEV}, I^{PV}, T^{Prompt}, G_{n-1}). \tag{1}$$

This formulation makes the previous-state map an explicit conditioning variable rather than a post-hoc stitching artifact, and it treats target-specific instructions as first-class inputs through text prompting (Yuan et al., 26 Sep 2025).

A plausible implication is that UniMapGen shifts map construction from a strictly perception-centric problem to a conditional generation problem in which map continuity, modality selection, and user intent can be handled in a single decoding process. The paper states this explicitly as a transition away from a local perception-plus-post-processing paradigm toward a more flexible, context-aware inference process (Yuan et al., 26 Sep 2025).

2. Sequence representation of lane maps

A central idea in UniMapGen is the conversion of vector maps into a discrete token sequence suitable for a decoder-only LLM. A map is formalized as

$$G = \{(\boldsymbol{v}_i, \boldsymbol{a}_i)\}_{i=1}^{K},$$

where $\boldsymbol{v}_i$ is a polyline and $\text{AP}^{M}=8.38$0 contains attributes such as category, line type, and start/end/cut type (Yuan et al., 26 Sep 2025).

Before serialization, each raw annotated polyline is resampled at a fixed spatial interval $\text{AP}^{M}=8.38$1, set to $\text{AP}^{M}=8.38$2:

$\text{AP}^{M}=8.38$3

This produces an equidistant representation $\text{AP}^{M}=8.38$4. The stated rationale is that long curves then receive more points, short segments receive fewer, and geometric spacing becomes regularized. This contrasts with fixed-point designs such as “20 points per polyline,” which the paper criticizes for geometric inefficiency (Yuan et al., 26 Sep 2025).

The framework then imposes a global ordering over lines. For each line, the first point’s distance to the origin is computed:

$\text{AP}^{M}=8.38$5

and lines are sorted in ascending $\text{AP}^{M}=8.38$6. The ordered map $\text{AP}^{M}=8.38$7 is serialized as

$\text{AP}^{M}=8.38$8

This converts the entire vector map into one sequence (Yuan et al., 26 Sep 2025).

The tokenizer is specialized. Numeric coordinates are encoded as single tokens such as \<257> rather than digit strings, structural field names such as points are also single tokens, and spaces are removed to control token count. The paper gives a JSON-like example in which a vector with coordinate list and attributes is converted into a compact token stream. The detokenizer inverses this process, parsing predicted tokens back into polylines and attributes (Yuan et al., 26 Sep 2025).

This sequence view is explicitly tied to autoregressive modeling. Although the paper notes that the training objective is standard next-token prediction, the generative behavior can be written as

$\text{AP}^{M}=8.38$9

This suggests that smooth continuation of lane geometry, category assignment, and boundary semantics are all learned in a single decoding stream rather than through a cascade of proposal, grouping, and refinement modules (Yuan et al., 26 Sep 2025).

3. Model architecture and multi-modal conditioning

UniMapGen is implemented as an MLLM-based generative framework using Qwen2.5-1.5B as the decoder-only transformer backbone (Yuan et al., 26 Sep 2025). The architecture combines modality-specific encoders with a joint token sequence fed into the LLM.

The BEV encoder is Dinov2 ViT-L/14, described as a 24-layer ViT with hidden size 1024 and 16 heads, producing BEV tokens $\text{AP}^{M}_{50}=21.67$0. The PV branch uses a 3D convolution over time followed by Qwen2-VL ViT, described as a 32-layer model with hidden size 1536 and 16 heads, producing $\text{AP}^{M}_{50}=21.67$1. Text is tokenized by Qwen2.5 into $\text{AP}^{M}_{50}=21.67$2, and the previous-state vector map is converted by the vector tokenizer into $\text{AP}^{M}_{50}=21.67$3. These are concatenated as

$\text{AP}^{M}_{50}=21.67$4

The model then autoregressively decodes the current map sequence conditioned on this joint input (Yuan et al., 26 Sep 2025).

The framework is designed to support multiple prompting modes. The supplementary prompt forms described in the data include BEV-only prompts such as “Please construct the entire road map in the satellite image,” PV-only prompts referring to perspective frames, and mixed prompts that specify coordinate ranges or trace points in the BEV image (Yuan et al., 26 Sep 2025). The same model therefore supports whole-map generation, target-road generation, and BEV–PV cooperative generation.

A key training mechanism is modality masking. During training, BEV, PV, text, or previous-map tokens are randomly masked, so the model learns to operate under different modality subsets. The stated combinations include BEV-only, PV-only, BEV+PV, BEV+Text, and BEV+PV+Text. This is presented as a robustness mechanism that enables dynamic selection among modalities at inference time (Yuan et al., 26 Sep 2025).

The reported optimization setup is standard language-model fine-tuning. Training uses AdamW with weight decay $\text{AP}^{M}_{50}=21.67$5, peak learning rate $\text{AP}^{M}_{50}=21.67$6, 6 epochs, batch size $\text{AP}^{M}_{50}=21.67$7, cosine learning rate decay with 100-step warm-up, and BF16 mixed precision with DeepSpeed (Yuan et al., 26 Sep 2025). The paper does not introduce a bespoke geometric loss at the decoder level; instead, smoothness and continuity are said to arise from the equidistant sampling, line reordering, and state-update conditioning.

4. State update and global continuity

A defining element of UniMapGen is its state update strategy for constructing large maps patch by patch. The global map after state $\text{AP}^{M}_{50}=21.67$8 is denoted $\text{AP}^{M}_{50}=21.67$9, with initialization

$=41.81$0

A large BEV image is partitioned into patches, and the model iterates over them in left-to-right, top-to-bottom order. At state $=41.81$1, the effective conditioning is

$=41.81$2

The previous map is therefore not merely auxiliary memory; it is part of the token input for current decoding (Yuan et al., 26 Sep 2025).

To make this workable, UniMapGen annotates line endpoints with three types: start, end, and cut. A cut endpoint denotes artificial truncation at a patch boundary. During training, these are derived from the ground-truth crop geometry. During inference, cut points from already generated patches are propagated into subsequent states and act as connection anchors for continuing lines across patches (Yuan et al., 26 Sep 2025).

The paper attributes cross-patch smoothness directly to this mechanism rather than to post-hoc optimization. Newly generated polylines are conditioned on adjacent cut points, so the model learns to continue existing roads rather than rediscover them independently in each patch. In the ablation study, enabling state update while keeping augmentation improves $=41.81$3 from $=41.81$4 to $=41.81$5 and mIoU from $=41.81$6 to $=41.81$7, and combining it with reordering and augmentation yields the full model’s $=41.81$8 and $=41.81$9, respectively (Yuan et al., 26 Sep 2025).

This suggests that UniMapGen’s continuity mechanism is not just spatial overlap but a learned causal relation between the already generated vector boundary and the next patch’s output. In that sense, the “state” is a compact symbolic map memory.

5. Handling occlusion, outdated imagery, and target-specific generation

UniMapGen is explicitly motivated by satellite imagery failure modes. For occluded roads, the paper states that the model uses visible BEV context before and after the occlusion, optional PV frames, and learned priors over lane continuity and topology. Because lane lines are generated as sequences of equidistant points, interpolation over hidden segments is learned as part of sequence prediction rather than introduced by a heuristic spline or graph completion stage (Yuan et al., 26 Sep 2025).

For outdated satellite imagery, the PV branch plays a corrective role. Ground-view images may show current lane markings that are worn out or absent in the satellite image. The paper’s qualitative example describes lanes marked in purple that are unclear in BEV but visible in PV, allowing UniMapGen to reconstruct up-to-date lanes from the joint prompt (Yuan et al., 26 Sep 2025).

Text prompts introduce a third mode of robustness: target-specific generation. When prompts specify coordinates or trace points, the model can generate only the requested road subset rather than the whole visible map. This is particularly important when annotation is incomplete. In the “Target GT” setting, the paper reports that BEV-only inference gives $n$0 and mIoU $n$1, while BEV+Text improves these to $n$2 and $n$3, and BEV+Text+PV further improves them to $n$4 and $n$5 (Yuan et al., 26 Sep 2025).

The claim that UniMapGen can infer occluded roads and predict roads missing from dataset annotations is part of the paper’s abstract and summary. The provided evidence is primarily qualitative plus the target-generation ablation. A plausible implication is that the framework learns a generative prior over road continuity strong enough to extrapolate beyond strict pixel evidence, though the paper also documents failure cases under extreme wear, severe occlusion without context, and limited prompt flexibility (Yuan et al., 26 Sep 2025).

6. Empirical performance, ablations, and adjacent research context

The principal quantitative benchmark is OpenSatMap20 validation. The paper compares UniMapGen with SegNeXt, MapTR$n$6, and MapTRv2$n$7. The reported numbers are as follows (Yuan et al., 26 Sep 2025):

Method	$n$8	$n$9	$$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$0	$$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$1	mIoU
SegNeXt	20.30	31.38	6.98	16.05	33.69
MapTR$$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$2	18.20	28.25	6.02	14.12	34.10
MapTRv2$$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$3	19.34	29.89	6.45	14.89	35.42
UniMapGen	29.17	34.81	8.38	21.67	41.81

The ablation study isolates three design factors: augmentation, line reordering, and state update. The baseline without any of the three achieves $$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$4 and mIoU $$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$5. Adding augmentation alone yields $$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$6 and $$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$7; adding reordering plus augmentation yields $$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$8 and $$G_n = \langle V, A \rangle = \{ (\boldsymbol{v}_i, \boldsymbol{a}_i) \}_{i=1}^K$$9; the full model with state update, reordering, and augmentation reaches $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$0 and $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$1 (Yuan et al., 26 Sep 2025). The paper therefore attributes substantial gains not only to the backbone and modality fusion but also to the serialization discipline itself.

On a nuScenes lane-topology evaluation, UniMapGen is reported against TopoNet, LaneGAP, and RNTR. The paper gives Landmark F1 $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$2 and Reachability F1 $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$3, with the latter exceeding RNTR’s $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$4 (Yuan et al., 26 Sep 2025). This indicates that the generated vectors preserve enough connectivity structure to improve graph-level navigation criteria, not only geometric overlap.

UniMapGen also sits within a broader research trend toward unified or generative map systems. UMPE, the “Unified Map Prior Encoder,” addresses mapping and planning by ingesting any subset of HD/SD vectors, rasterized SD maps, and satellite imagery, using alignment-aware vector and raster branches; on nuScenes mapping, it lifts MapTRv2 from $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$5 to $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$6 mAP and reduces VAD planning error from $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$7 to $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$8 m average L2 (Zhang et al., 4 May 2026). GenMapping pursues sensor-agnostic online HD map generation through inverse perspective mapping and reports cross-dataset generalization gains over conventional camera-dependent BEV pipelines (Li et al., 2024). MapDreamer formulates lane-level map generation from aerial imagery as latent diffusion over vector lane graphs with explicit topology, reporting stronger GEO and TOPO F1 than BGFormer on UrbanLaneGraph-derived benchmarks (Brandes et al., 1 Jul 2026). These works define adjacent but distinct design directions: UniMapGen is sequence-generative and multi-modal at the token level; UMPE is a unified prior encoder; GenMapping is an IPM-centric universal online mapper; MapDreamer is a conditional generative vector diffusion model from aerial imagery.

7. Limitations and significance

The paper identifies several limitations. Inference is relatively heavy: approximately 4 seconds per $\boldsymbol{v}_i = \{(x_j, y_j)\}_{j=1}^{N_j}.$9 patch on GPU, making the system more suitable for offline large-scale mapping and updates than for real-time onboard deployment (Yuan et al., 26 Sep 2025). Text prompting is also narrow: current prompts are largely coordinate-based or trace-based rather than free-form natural-language geographic instructions. The paper states that prompts like “generate map from the school to the hospital” are not supported because of missing training data and landmark annotations (Yuan et al., 26 Sep 2025).

Failure cases include extreme wear or outdatedness without PV context and severe occlusion without enough surrounding cues. In such cases, the model may miss or misplace lane lines. These are consistent with the framework’s generative bias: when pixel evidence and prior context are both weak, the sequence model may not have enough grounding to produce the correct geometry (Yuan et al., 26 Sep 2025).

Despite these limits, UniMapGen is significant because it consolidates several previously separate ideas into one framework: vector map serialization, multi-modal prompting, and stateful global continuity. The framework’s reported improvement over strong satellite baselines on OpenSatMap20, together with its ability to exploit PV and text for occlusion recovery and target-road generation, positions it as a representative example of large-scale map construction treated as conditional generative modeling rather than local vector prediction (Yuan et al., 26 Sep 2025).

A plausible broader implication is that UniMapGen defines a useful abstraction for future mapping systems: map construction as token generation over structured vectors, conditioned on heterogeneous sensory and symbolic context. In adjacent literature, unified prior integration (Zhang et al., 4 May 2026), sensor-agnostic online mapping (Li et al., 2024), and generative lane-graph synthesis (Brandes et al., 1 Jul 2026) suggest that the field is converging on unified representations that combine geometry, topology, appearance, and prior knowledge. UniMapGen’s specific contribution within that trend is the claim that large-scale map construction can be made sequence-native, multi-modal, and stateful in a single decoder-driven framework (Yuan et al., 26 Sep 2025).