acceptodds
Under review as a conference paper at ICLR 2027

CityExpander: Agent-Guided Synchronized Diffusion for Large-Scale City Generation

Abstract

Generating large-scale urban semantic layouts from text requires preserving both city-wide spatial coherence and fine-scale structural detail while satisfying spatial planning constraints. Directly generating high-resolution semantic image over large urban areas is computationally expensive, while lowering the resolution can erase narrow roads and merge nearby buildings. Sliding-window generation limits per-window computation but introduces sequential dependencies and may produce seams or repetitive patterns. Meanwhile, directly conditioning a generative model on text provides only coarse control over the locations, proportions, and spatial relations of land-use categories. We present CityExpander, a text-guided diffusion framework that addresses both challenges. First, We introduce a compiler-grounded condition plan agent that translates natural-language descriptions into executable zone-ratio conditions. To balance computational cost and structural fidelity, we use a global DiT for coarse city structure and a local DiT for fine-scale refinement. We propose semantic consensus globally synchronized denoising (SC-GSD) to achieve semantic consistent layout generation with simultaneous denoising, avoiding the sequential dependency of sliding-window sampling. Extensive experiments demonstrate that CityExpander effectively generates high-quality text-guided city layouts at both single-tile and large spatial scales. Further evaluations validate the effectiveness of SC-GSD in promoting semantic consistency and reducing time cost during large-scale synthesis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
City Layout Generation, Synchronized Diffusion

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_d4oyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CityExpander for generating large-scale city layouts from text. The main idea is to first use an LLM to turn the text description into zoning and land-use constraints, and then generate the layout in two stages: a global stage for the overall structure and a local stage for details. The authors also introduce SC-GSD to reduce inconsistencies between overlapping local regions. The experiments are conducted on data derived from OSM, and the results show that CityExpander generally performs better than the compared methods in terms of layout quality, controllability, and generation efficiency.

Critical strengths
  1. The paper clearly identifies the key challenges in large-scale city layout generation: maintaining global coherence, preserving local geometric detail, and enabling precise spatial control from text. The motivation is convincing and well connected to the limitations of existing sliding-window and direct text-conditioning approaches. Based on it, CityExpander combines a compiler-grounded planning agent, a global–local diffusion pipeline, and semantic-consensus synchronized denoising in a natural way. The intermediate zoning-ratio representation is sensible, as it provides a practical bridge between high-level language instructions and low-level spatial generation.
  2. The experiments cover single-tile quality, text-guided generation, large-area synthesis, computational efficiency, and several component ablations, demonstrating its capability.
Critical weaknesses

The evaluation of text–layout alignment could be more direct. Table 2 measures similarity between generated and reference layouts using CLIP features and category-ratio errors, but neither metric explicitly evaluates whether the generated layout follows the input description. A text–image similarity metric would therefore be a useful addition. Additionally, the paper highlights directional control—for example, placing an urban area in the east—as an advantage over direct text conditioning, yet the evidence for this claim is mainly qualitative. A small targeted evaluation covering directional and regional constraints would make the controllability claim substantially more convincing.

Questions for authors
  1. Please See Critical Weaknesses.
  2. The global stage is directly constrained by the zoning-ratio field, while the local stage appears to depend mainly on the coarse structural latent. It is unclear whether this latent alone is sufficient to preserve the intended regional composition throughout local refinement. Some additional analysis or ablation on this point would be useful.
  3. It would also be useful to report the end-to-end inference time of CityExpander and the baselines, including the time required by the Condition Plan Agent. Furthermore, a comparison with MajutsuCity (CVPR26) would strengthen the evaluation, as it addresses a closely related language-driven text-to-layout setting.
  4. Section 3.1 assigns an intended function s_k to each region, and the planner is described as organizing functional zones with explicit land-use compositions. As I understand it, only the compiled class-ratio field is passed to the diffusion model. This means that, for example, residential and commercial regions with the same geometry and class ratios would result in identical conditioning, despite having different functional labels. Since these labels may be useful for practical personalization, the authors could consider incorporating them more directly into the generation process to better preserve the functional distinctions specified by the user.
2 comments
AuthorsOpenReview ↗

Response to Reviewer d4oy (Part 1/2)

We thank the reviewer for the careful reading and constructive feedback.

  1. Direct evaluation against the input text. We thank the reviewer for suggesting a direct evaluation against the input text and quantitative evidence of directional and regional control. We used GPT-6-astra to derive 4,353 programmatic unit tests and 4,908 open-ended semantic tests from the 1k-text prompts. The programmatic tests automatically check measurable constraints on the generated semantic layouts, including class presence, regional density comparisons, and regional concentration, with rules and thresholds fixed before scoring. The open-ended tests use VLM-assisted visual assessment to evaluate requirements involving shapes, spatial relations, topology, and qualitative attributes. A requirement passes when a reasonable interpretation is supported by visible evidence. GPT-6-Astra is used as judge. We report constraint satisfcation rate (CSR) here.

    MethodCLIP-sim ↑Cat. L1 ↓Cat. JS ↓CLIP-T ↑Rule CSR ↑Open CSR ↑Overall ↑
    CityCraft-Text0.8950.3150.0390.26689.3644.0565.35
    Ours (Qwen)0.9080.2550.0270.26598.2284.5691.00
    Ours (Gemini)0.9110.2370.0200.26396.5378.1686.79

    The pass rates for selected evaluations of directional and regional control are reported below:

    EvaluationNumber of testsCityCraft-TextCityExpander(Gemini)CityExpander(Qwen)
    Class presence in a specified half or quadrant ↑196287.87%95.87%98.42%
    Relative density between specified regions ↑13150.38%94.66%93.89%
    Concentration in a specified region ↑3141.94%83.87%74.19%

    The results show that CityExpander achieves better instruction-level semantic alignment than CityCraft-Text with both Gemini and Qwen. The benchmark already includes subsets for regional and directional instructions, where CityExpander also outperforms CityCraft-Text by a large margin. These results further support CityExpander’s stronger text–layout alignment.

    CLIP-T (CLIP text-image similarity) slightly favors CityCraft-Text despite its lower constraint satisfaction rates. We interpret this discrepancy as a failure of CLIP-T to reflect the spatial compliance improvements measured here. A plausible explanation is that global image–text similarity emphasizes shared semantic content without sufficiently distinguishing spatial arrangements: a layout may contain the requested semantic classes while placing them in incorrect regions or failing to satisfy their relative-density requirements. Consequently, CLIP-T alone is insufficient to evaluate instruction-level spatial control in this setting.

  2. We thank the reviewer for highlighting the need to measure composition changes during local refinement. We compared the Global and Final outputs against the same compiled zoning-ratio field, aggregating each layout onto the condition grid and measuring mean absolute error over the vegetation, building, road, and water channels. On the 25 regular prompt families across four scales, the SC-GSD results are:

    Class or aggregateMAE on Global StageMAE on Final Image
    Vegetation0.1910150.188626
    Building0.1429700.139293
    Road0.0783220.084611
    Water0.0579420.058347
    Mean over four channels0.1175620.117719

    The mean error increases by 0.000157 in ratio units. This analysis shows limited additional deviation from the conditioning field during Local refinement on the evaluated cases.

AuthorsOpenReview ↗

Response to Reviewer d4oy (Part 2/2)

Q3. End-to-End Generation Cost. We measured end-to-end time on the same 25 prompts using one RTX 4090. Times include planning, sampling, decoding, and saving, but exclude model loading and warmup.

MethodPlanning time (s)Total time (s)
CityCraft-Text-4.95
CityExpander96.52114.14

We report end-to-end time only for single-patch generation because CityCraft-Text only supports text-conditioned generation of a fixed-size patch. Extending it to larger scenes would also require spatial planning to coordinate multiple patches. Our agent planning and sampling together take longer than the baseline. This extra time is expected because planning involves several remote LLM calls to generate and revise the plan. This extra cost supports a more capable generation pipeline. Our experiments show better layout quality with ratio conditioning and stronger instruction-level semantic control with agent-planned conditions. The agent also separates the text description from the output scale, allowing the same description to guide large-scale scene generation. In contrast, the text-condition variation of CityCraft baseline maps each description to a fixed-size patch. The comparison with MajutsuCity requires retraining the baseline, and the experiments have not been done due to limited computational resources. We will share the results if the experiments are completed.

Q4. Semantic label degradation. We agree with the reviewer that incorporating functional labels more directly into generation could better support personalization. Our original motivation for introducing semantic labels was to provide an intermediate planning step inspired by chain-of-thought reasoning. These labels guide and constrain the assignment of class ratios, rather than asking the agent to directly predict numerical ratios for all five semantic classes.

With the current backbone unchanged, one possible approach is to assign finer-grained functions downstream. The original region labels and user instructions can be retained as metadata and provided alongside the generated layout to an LLM-based planner. CityCraft demonstrates this separation: an LLM assigns building functions and architectural attributes after layout generation, followed by asset retrieval and 3D scene construction.

Another approach is to incorporate functional labels directly by expanding the land-use categories generated by the backbone, allowing it to output finer-grained semantic classes such as residential and commercial areas. We will explore these directions in future work.

Reviewer_PRuDOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

CityExpander studies large-scale 2D urban semantic layout generation from natural-language descriptions. Its main contribution is to convert input text prompts into an explicit intermediate representation, a spatial zone-ratio conditions, which guides subsequent layout generation. The framework also tailors its diffusion pipeline to urban layout generation, using global–local diffusion and SC-GSD to fuse predictions from overlapping windows and improve generation quality.

Critical strengths
  1. The paper introduces an intermediate representation that extends the global ratio vector used in CityCraft. The spatial zoning-ratio field explicitly represents the land-use compositions and locations of different regions, providing a direct means of controlling localized spatial requirements expressed in text.

  2. The authors tailor the diffusion architecture and interactions among sampling windows to urban layout generation, and experimentally demonstrate the effectiveness of these modifications.

Critical weaknesses
  1. A core contribution of the paper is to use the Condition Plan Agent and spatial zoning-ratio fields to translate textual requirements into generation conditions with explicit locations and spatial relationships. However, the experiments lack an independent quantitative assessment of whether these fields satisfy the spatial constraints in the original text.

  2. The improvement offered by spatial zoning-ratio fields over global ratio vectors has not been sufficiently established. In Table 4 of the appendix, CityCraft receives the area-weighted average ratios computed from the same zoning plan, giving both methods a common planning source. However, this comparison changes both the conditioning representation and the generative model: CityCraft uses a global ratio vector, whereas CityExpander uses a spatial zoning-ratio field together with its own generation framework. The resulting quality improvements therefore cannot be attributed solely to the spatial representation.

Questions for authors
  1. What level of descriptive granularity can the Condition Plan Agent handle? Most examples in the paper focus on explicit spatial relationships among urban elements and achieve good results. I am particularly interested in whether the Condition Plan Agent can interpret descriptions of urban style and translate them into spatial zoning-ratio fields.
  2. How is the spatial resolution of the condition cells produced by the Condition Plan Agent determined? How would increasing or decreasing the condition-grid resolution affect spatial constraint satisfaction and generation quality?
  3. The spatial zoning-ratio field can be viewed as an image-like layout representation and is already rendered as a preview for VLM-based review within the Condition Plan Agent. Could this preview be used to condition layout generation models such as CityGen and CityDreamer? This could help assess the generalizability of the spatial zoning-ratio representation and provide an additional baseline for a fair comparison with CityExpander.
  4. In the w/o Global ablation, is the spatial zoning-ratio input also removed? The ratio field appears to have the same grid dimensions and physical spatial resolution as the global latent. Could it be used directly as a conditioning input to local diffusion to provide a more controlled ablation? This experiment could help determine whether the spatial zoning-ratio field already provides sufficient global information and how much additional improvement in generation quality is contributed by global diffusion.
4 comments
AuthorsOpenReview ↗

Response to PRuD (Part 1/2)

We thank the reviewer for the careful reading and constructive feedback.

W1. Evaluations on zone-ratio field. We thank the reviewer for highlighting the need to evaluate the intermediate zoning-ratio field independently of layout generation. On 25 prompt family over 1-4km (100 cases), we have added a direct evaluation of the compiled fields against constraints extracted from the original prompts. Based on the original text, we developed a rule-based evaluation procedure to assess the generated zone ratios against 304 verifiable constraints and compute the constraint satisfaction rate. The results show that all evaluated cases satisfied all applicable constraints, demonstrating the robustness of Condition Plan Agent.

W2. Ratio conditional generation. We agree that the original comparison does not isolate the representation alone. We revised the comparison to include two variants of CityCraft: one conditioned on a ratio vector(CityCraft-RV) and the other on a ratio field (CityCraft-RF). The latter corresponds to our CityExpander Global variant. This setup enables a fair comparison.

We conducted separate comparisons using ratio fields computed from the test set and agent-planned ratio fields derived from the 1k-text test set.

We use average category error( ACE) and zone-average category error (zACE) introduced in appendix B to compare the semantic-class area proportions in generated layouts with those specified in the target plans.

Condition sourceNMethodFID ↓KID ↓CMMD ↓ACE ↓zACE ↓
Agent-planned1000CityCraft-RV76.13360.04003417.52580.0088240.128742
Agent-planned1000CityCraft-RF63.00330.02332515.83370.0264550.097364
Test-Driven10000CityCraft-RV46.26060.02963015.61850.0065650.107453
Test-Driven10000CityCraft-RF42.10600.02533917.64340.0221370.086730

Across both condition sources, CityCraft-RF consistently achieves lower FID, KID, CMMD, and ACE than CityCraft-RV, indicating improved generation quality and better alignment with the target spatial distributions. CityCraft-RV’s lower ACE is expected, as it is directly conditioned on the global ratio vector during training. In contrast, local deviations from the ratio-field conditions in CityCraft-RF can collectively lead to a larger global ratio error. Nevertheless, this error remains small in absolute terms, with ACE values of 0.0265 and 0.0221 under agent-planned and test-driven conditioning, respectively.

Q1. Descriptive granularity. We thank the reviewer for asking about the descriptive granularity of the Condition Plan Agent. Given an urban-style description, the agent can interpret it in terms of spatial organization, functional zones, and regional class proportions, and translate these attributes into a spatial zoning-ratio field. For example, given “generate a cyberpunk district,” the agent can plan a commercial core, dense residential areas, night markets, and industrial workshops, specifying their locations, shapes, and semantic-class area proportions to guide layout generation.

However, appearance attributes, including building details, materials, and architectural styles, are not encoded in this ratio field. Differences in these attributes therefore cannot be conveyed to the generation model through the field or enforced as layout constraints. This representation supports translating style descriptions into spatial layout characteristics, while control over specific architectural appearance requires additional mechanisms in downstream generation.

AuthorsOpenReview ↗

Response to PRuD (Part 2/2)

Q2. Resolution of condition cell. The condition-cell resolution is a fixed design choice aligned with the spatial resolution of the global diffusion latent. The Condition Plan Agent specifies regions and their land-use compositions, and the deterministic compiler converts this plan into a ratio field on the predefined grid. For larger scenes, the grid expands while the physical size of each cell remains fixed, maintaining a consistent spatial meaning for the conditions across scales. We will clarify this setting in the revised manuscript.

Ablations on resolution of condition cell. We evaluated five effective condition-grid resolutions on 25 instruction families at 1 km, using three matched seeds per family (75 outputs per resolution). Following the previous approach, we utilized GPT-6-astra to convert the 75 text descriptions into 231 programmatic evaluation rules and 216 open-semantic evaluation criteria. The corresponding pass rates are reported below. We also report ACE and zACE.

Effective gridRule CSR ↑ (%)Open-Sem. CSR ↑ (%)zACE ↓ (x100)ACE ↓ (x100)
1×195.6765.2813.6025.109
4×498.2773.6112.6974.668
8×897.4079.1712.3774.455
16×1698.2781.0212.2614.379
32×3297.8480.0912.0784.211

Overall, CSR increases as the cell size decreases, while both ACE metrics decrease consistently. This indicates that finer cell partitioning can improve the constraint satisfaction rate and reduce discrepancies in zone ratios between generated layouts and initial plans.

Q3. Ratio field. Our comparison of ratio-vector and spatial ratio-map in W2 demonstrates applicability to the DiT backbone used by CityCraft, which provide partial evidence on the generalizability of the zone ratio field. Due to limited computational resource, adaptation on CityGen and CityDreamer have not been done yet. We will report their results if completed

Q4. Ablation on Global. We thank the reviewer for this suggestion. The Table 1 w/o Global ablation belongs to the unconditional setting; its two-stage baseline also receives no ratio field. We additionally fine-tuned a Local model directly conditioned on the spatial ratio field for 100k steps and evaluated 1,000 cases against Full, using identical ratio fields, GT, and 20-step DDPM Local sampling. The results provides strong evidence of the importance of Global model.

MethodFID ↓KID ↓
Ratio-conditioned Local (100k fine-tuning)253.9280.222963
Full (Global + Local)64.8790.021406

Thank you for the clarification and response. I have no further questions, and I will maintain my positive score.

Thanks for your response!

Reviewer_CTmNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CityExpander, a text-guided diffusion framework for generating large-scale urban semantic layouts. The method has three main components: (1) a condition planning agent that translates natural-language descriptions into hierarchical spatial regions and land-use ratio maps using an LLM and a deterministic compiler; (2) a global–local diffusion pipeline that generates coarse city structures and refines them into high-resolution layouts; and (3) Semantic-Consensus Globally Synchronized Denoising (SC-GSD), an inference-time strategy that synchronizes overlapping diffusion windows using noise fusion and semantic consistency constraints. Experiments on OpenStreetMap-derived layouts evaluate single-tile generation quality, text-to-layout alignment, and large-area seam consistency. The results suggest improvements over several baselines in layout quality and spatial controllability, while procedural 3D city generation demonstrates a potential downstream application.

Critical strengths
  1. The paper addresses two important challenges in large-scale urban layout generation: maintaining global spatial coherence while preserving fine-grained structures, and translating high-level textual descriptions into localized spatial constraints.

  2. Extensive experiments have been conducted to verify the effectiveness. Beyond image-generation metrics, the paper evaluates road connectivity, road orientation, building-area distributions, category ratios, and cross-window consistency. The component ablations also help illustrate the respective roles of global generation and local refinement.

  3. This paper is overall well-written and easy to follow.

Critical weaknesses
  1. The proposed method builds upon shared-canvas multi-window sampling and noise fusion, as in MultiDiffusion and SyncDiffusion, and introduces semantic consistency losses during inference. While the additional constraints are intuitively reasonable, the methodological novelty appears incremental.

  2. Table 1 states that the full model uses test-derived zoning-ratio conditions, whereas the unconditional variant uses an all-zero condition. Since the conditions for the full model are derived from the test data, the reported improvements may partly reflect access to ground-truth spatial statistics rather than better unconditional generation. The paper should clearly distinguish unconditional generation, generation under oracle conditions, and generation under conditions produced by the proposed planning agent. These settings should be evaluated separately, with matched conditions wherever possible. It is also important to clarify whether the text-to-layout evaluation uses independently specified constraints or information derived from the reference layouts.

  3. The main large-area sampler comparison evaluates maps spanning 2.5km × 2.5km, while the planning agent is tested at scales up to 4km. This provides limited evidence for the claimed scalability to large urban areas. The paper would benefit from experiments at larger spatial extents, reporting runtime, peak memory consumption, seam consistency, and road connectivity as the number of windows increases.

  4. The reported 100% final success rate is based on compiler-valid outputs after retries and bounded repairs. This demonstrates robustness of the execution pipeline but does not establish that the resulting plans correctly interpret natural-language instructions. The paper should report the semantic correctness of generated plans, the accuracy of localized spatial constraints, and the frequency of invalid or semantically incorrect plans before repair.

Questions for authors
  1. How are the test-derived zoning-ratio conditions in Table 1 constructed? Are they computed from ground-truth test layouts, and can the authors provide a comparison using only independently specified or agent-generated conditions?

  2. Can the authors evaluate larger spatial extents and report how runtime, memory, and structural consistency scale with the number of overlapping windows?

  3. Can the authors provide stronger evidence of road-network realism and instruction-level spatial controllability beyond image-based metrics and aggregate category statistics?

6 comments
AuthorsOpenReview ↗

Response to Reviewer CTmN (Part 1/2)

We thank the reviewer for the careful reading and constructive feedback.

  1. Contributions. We thank the reviewer for raising this point. SC-GSD builds on the shared-canvas, multi-window denoising principle of prior work. Its contribution is to incorporate semantic-consensus constraints tailored to urban layouts, addressing the trade-off between structural consistency, generation quality, and runtime as the generated area increases. Besides, CityExpander also contributes a spatial zoning-ratio representation with a Condition Plan Agent, which translates textual requirements into localized spatial conditions, and a global–local diffusion pipeline, which separates city-scale structure generation from local detail refinement. The representation and planning agent enable finer spatial control from text, while SC-GSD and multi-scale diffusion address complementary aspects of consistency, quality, and efficiency in large-area generation. Together, these components form a coordinated framework for controllable city-layout synthesis.

  2. Test-derived zoning-ratio conditions. We thank the reviewer for highlighting the need to distinguish the conditioning settings. The full-model versus all-zero-condition comparison in Table 1 is intended to evaluate whether the trained ControlNet can exploit previously unseen zoning-ratio maps to improve generation quality. These conditions are computed by aggregating semantic-class area proportions over spatial regions of held-out ground-truth test layouts. The comparison measures the benefit of ratio conditioning on held-out samples, rather than establishing better unconditional generation.

    End-to-end text-layout generation on 1k test samples is evaluated separately in the quantitative comparisons in Table 2, where we add extra metrics on text-image in the revised manuscript. And we would like to clarify that the text prompt in Table 2's text-to-layout evaluation were annotated from test split by Gemini-2.5-Pro and validated by human. Text prompts on robustness evaluation were independently specified.

To enable fair comparison among uncond., full(oracle) and agent-generated scenes, we evaluated 1,000 scenes sampled using these methods. The results below demonstrate that agent-planned conditions also improve the quality of generated urban scenes, although the gains are smaller than those achieved with oracle conditions. The results will be also added in (Appendix C) of the revised version.

MethodFiDKIDHICMMDRoad L1Road JSAng. EMDBld. EMD
Uncond.77.260.0330.69615.870.10220.00260.93126.28
Full (Planned)72.280.0320.65517.630.07860.00181.04325.29
Full (Test-Drivem)61.360.0170.65811.350.06930.00110.4587.88
  1. Evidence for scaling. We thank the reviewer for suggesting a more systematic scalability evaluation. We have completed generation experiments at lengths of 1,2,3,4 and 8 km, with 25 prompt families per scale . We report the completed results for our SC-GSD pipeline below. The results will be also added in (Appendix C) of the revised version.

    ScaleGlobal / Local windowsSampling time (s)Global peak memory (GiB)Normalized seam MSERoad-crossing score (%)Road LCC (%)
    1 km1 / 4918.42.260.97588.5464.22
    2 km9 / 22565.02.711.06486.1266.36
    3 km25 / 529153.43.691.02785.5458.16
    4 km49 / 961305.55.061.03385.7467.24
    8km225/39691185.314.471.03184.3564.48

    Sampling time sums Global and Local sampling, excluding planning, model loading, and final decoding; memory denotes the peak allocated GPU memory. Structural metrics are measured on the final layouts: normalized seam MSE compares Local-window boundaries with non-boundary control locations, the road-crossing score measures adjacent-boundary road alignment with a 2 m tolerance, and LCC is the fraction of original road pixels in the largest connected component.

    Metrics on show no degradation as the generation scale increases, providing further evidence of scalability.

AuthorsOpenReview ↗

Response to Reviewer CTmN (Part 2/2)

  1. Semantic correctness of planned field. We agree that compiler validity alone does not establish semantic correctness. We have therefore added a constraint-based evaluation on 100 cases from 25 prompt families, covering regional composition and localized spatial requirements. The planned ratio fields and final layouts achieve 100% and 99.7% satisfaction rates, respectively, on the annotated constraints.

  2. Evidence of road structure and instruction-level spatial control. We thank the reviewer for requesting stronger evidence of road structure and instruction-level spatial control. Aggregate class proportions provide coverage checks; we additionally evaluated road-skeleton structure and explicit spatial requirements.

    On the paired 1,000-text benchmark, the distances between generated and reference road-skeleton degree distributions are lower for ours: RoadL1 is 0.078644 versus 0.149045 for CityCraft-Text, and RoadJS is 0.001822 versus 0.005669.

    We also compared SC-GSD, vanilla GSD, and sliding-window sampling on 100 paired cases at the nominal 2.5 km scale, holding local refinement fixed within each comparison. Road-skeleton endpoint density within eight pixels of the inherited global seams is:

    Fixed local refinementSC-GSDVanilla GSDSliding
    SC-GSD Local0.0340290.0365540.035065
    Vanilla GSD Local0.0449040.0486830.045554

    Endpoint density is the fraction of skeleton pixels with one eight-connected neighbor. SC-GSD has the lowest seam-neighborhood density under both refinement settings; Sliding has lower whole-image endpoint density. Thus, the evidence supports improved continuity near seams, with natural road ends and cropping boundaries remaining contributors to this proxy. These results are added in Appendix C revised version.

    For instruction-level control, we used GPT-6-astra to derive 4,353 programmatic unit tests and 4,908 open-ended semantic tests from the 1k-text prompts. The programmatic tests automatically check measurable constraints on the generated semantic layouts, including class presence, regional density comparisons, and regional concentration, with rules and thresholds fixed before scoring. The open-ended tests use VLM-assisted visual assessment to evaluate requirements involving shapes, spatial relations, topology, and qualitative attributes. A requirement passes when a reasonable interpretation is supported by visible evidence. GPT-6-Astra is used as judge. We report constraint satisfcation rate (CSR) here.

    MethodCLIP-sim ↑Cat. L1 ↓Cat. JS ↓CLIP-T ↑Rule CSR ↑Open CSR ↑Overall ↑
    CityCraft-Text0.8950.3150.0390.26689.3644.0565.35
    Ours (Qwen)0.9080.2550.0270.26598.2284.5691.00
    Ours (Gemini)0.9110.2370.0200.26396.5378.1686.79

    The pass rates for selected evaluations of directional and regional control are reported below:

    EvaluationNumber of testsCityCraft-TextCityExpander(Gemini)CityExpander(Qwen)
    Class presence in a specified half or quadrant ↑196287.87%95.87%98.42%
    Relative density between specified regions ↑13150.38%94.66%93.89%
    Concentration in a specified region ↑3141.94%83.87%74.19%

    The results show that CityExpander achieves better instruction-level semantic alignment than CityCraft-Text with both Gemini and Qwen. The benchmark already includes subsets for regional and directional instructions, where CityExpander also outperforms CityCraft-Text by a large margin. These results further support CityExpander’s stronger text–layout alignment. CLIP-T (CLIP text-image similarity) slightly favors CityCraft-Text despite its lower constraint satisfaction rates. We interpret this discrepancy as a failure of CLIP-T to reflect the spatial compliance improvements measured here. A plausible explanation is that global image–text similarity emphasizes shared semantic content without sufficiently distinguishing spatial arrangements: a layout may contain the requested semantic classes while placing them in incorrect regions or failing to satisfy their relative-density requirements. Consequently, CLIP-T alone is insufficient to evaluate instruction-level spatial control in this setting.

AuthorsOpenReview ↗

Additional reply for Q3

Additional evidence on scability.

In 8km generation on 25 prompt family, we compare different sampling strategies on global stage. The results is shown on the table below.

MethodSampling Times (s)Peak GPU Memory (GiB)LCC (%) ↑LCC@8 m (%) ↑Endpoint Density ↓Road L1 ↓Road JS ↓Road F1 ↑
SC-GSD750.314.470.239326.79224.610.154630.00612076.54
Vanilla GSD650.513.680.218618.95229.110.162780.00684974.53
Sliding2527.013.680.222220.94227.260.157610.00614676.56

SC-GSD beats vanilla GSD on all metrics and beats sliding on most of metrics, demonstrating its advantage in semantic continuity. It also substantially reduces sampling time compared with sliding-window sampling, while incurring only modest runtime overhead over vanilla GSD and a slight increase in GPU memory usage over the baselines. The results provide more evidence on advantage of our proposed SC-CSD on large scales.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised Version Uploaded

A revised version of our paper has been uploaded. All changes are highlighted in green. In the main text, we have added clarifications and experiments evaluating instruction-level spatial controllability using ruled-based unit test and VLM judge, which may help address the reviewers’ concerns about text–layout alignment (CTmN and d4oy) and the semantic correctness of agent-generated plans (CTmN and PRuD).

In the appendix, we have included supplementary experiments on the effectiveness and generalizability of the ratio representation (PRuD), comparisons between unconditional oracle ratio-conditioned and agent planned ratio-conditioned generation (CTmN), as well as the end-to-end generation time for a single tile (d4oy), providing a more comprehensive evaluation of our method. Other questions are replied in the comments.

We hope that our replies and revised manuscript address the reviewers’ concerns. Please let us know if any questions remain; we would be happy to respond.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.