Greenscapes: Optimized Vegetation Configurations in Cities

2 Related Work

• Existing urban vegetation methods often simplify planting to stochastic points or trees, while generative approaches generally do not couple landscape rules with quantifiable 3D performance metrics.The paper identifies limited controllability over setbacks and functional goals as another gap.Source: §1 · pp. 1–2

1 Introduction

• The framework couples VLM structural priors with CMA-ES geometric refinement, evaluating shade, sky-view factors, acoustic attenuation, and view-dependent aesthetics.It supports heterogeneous vegetation and design typologies while adhering to parcel geometry.Source: §1 · p. 2

3 Method

• The pipeline uses a text prompt and OSM lot geometry to generate top-down proposals, extract vegetation instances, assemble a 3D scene, evaluate it, and return the best feasible configuration.Its experiments use fixed prompt templates, and the scene is refined using CMA-ES with a GMM asset prior.Source: §3 · pp. 2–3

3.1 OSM Preprocessing, Masks, and Assets

• OSM geometry is rasterized into aligned masks that distinguish exclusion and mandatory regions, while parameterized large-tree, small-tree, and shrub assets carry geometric and functional attributes.Asset attributes include ellipsoid bounds and coefficients for shade, biomass, and acoustic damping.Source: §3.1 · p. 3

3.2 Top-down Layout Generation and Instance Extraction

• The constrained generator edits grass pixels and proposes vegetation layouts; a second VLM pass extracts labeled instances, which are deterministically repaired or discarded if they remain infeasible.Repair moves centers to feasible planting locations and shifts ellipsoids away from remaining obstacles or overlaps.Source: §3.2 · pp. 3–4

3.3 In-scene Rendering and Metric Computation

• Rendered views supply skyline and sky-view metrics, while biomass and efficient shade and acoustic surrogates evaluate candidate scenes from selected viewpoints.The authors explicitly characterize the functional terms as design surrogates rather than calibrated physical predictions.Source: §3.3 · p. 4

3.4 CMA-ES Refinement

• After one feasibility repair, CMA-ES keeps instance positions fixed and optimizes per-instance scales and attribute proposals, with GMM assignments enabling discrete species substitutions.Candidate configurations are evaluated using weighted SVF, skyline, path-shading, acoustic-exposure, and biomass terms, plus a scale regularizer.Source: §3.4 · p. 5

Park-Design Strategies

• Orientation entropy and linearity separate three design strategies, and generated parks align with corresponding reference distributions, supporting metric validity and style control.The compared strategies are formal, picturesque, and modernist.Source: §4 · p. 6

Prompt-conditioned Controllability

• Changing prompts for one park produces plausible alternatives, from haphazard agroecological planting to sparser, more organized modernist and memorial layouts.The layouts are compared with a reconstructed St. Mary’s Park baseline.Source: §4 · p. 6

CMA-ES Refinement

• With planting positions fixed, CMA-ES changes perceived design through rescaling and asset substitutions, shaping skyline, openness, shade, acoustic screening, or biomass.The skyline-targeted run produced alternating canopy heights; increasing SVF favored slimmer crowns, while maximizing biomass favored broader canopies and higher scales.Source: §4 · p. 6

Planting Layouts of Lots

• The prompt-conditioned pipeline qualitatively reproduces semi-random, boundary-biased, and clustered planting patterns from Niese et al. [2022].Source: §4 · p. 6
• Across varied OSM lots, generated planting remains constraint-adherent and adapts to remaining ground area, filling edges, setbacks, corners, wedges, and narrow corridors.Source: §4 · p. 6

Quantitative Validation and Robustness

• Across 30 runs, CMA-ES reduces mean target error and run-to-run variance versus the VLM prior, improving objective alignment (Table 1).The VLM prior is sample-sensitive, including in skyline-related error.Source: Table 1 · p. 7, §4 · p. 7

Runtime Analysis

• A 100-generation experimental run takes approximately 38 minutes on the reported laptop, with VLM generation and detection adding a one-time 15–20-second initialization.One optimization generation averages approximately 23 seconds on that system.Source: §4 · p. 7

Quantitative Validation and Robustness

• The VLM prior improves shade over random placement, and VLM-initialized CMA-ES achieves the highest shade and acoustic scores while minimizing biomass.Random-start CMA-ES improves functional scores but is less effective than VLM-initialized refinement because random layouts often place vegetation away from paths.Source: §4 · p. 7

5 Discussion and Limitations

• Cost is 0.006 for CMA-ES on the VLM prior, versus a random initialization that ends about 20× farther from the joint shade and biomass target.The VLM-prior result reaches shade 2.04 against a 2.00 target and keeps biomass within about 2.3% of baseline; random initialization reaches shade 1.91 and overshoots biomass by about 5.5%.Source: Table 2 · p. 7
• Unlike prior work focused primarily on trees, GreenScapes configures heterogeneous plantings under coupled functional and perceptual objectives.The cited comparison names Niese et al. [2022] as focusing primarily on trees and not explicitly coupling functional performance with perceptual architectural strategies.Source: §5 · p. 7
• CMA-ES refines a low-dimensional 3D parameterization against scene-derived objectives, while the image-edit prior supplies stylistically coherent, human-plausible structure.This division addresses the difficulty of formalizing qualitative landscape-architectural principles procedurally.Source: §5 · p. 7
• The functional metrics are approximate optimization surrogates rather than first-principles physical simulations, and the method runs slower than procedural vegetation configuration approaches.The discussion contrasts its runtime with Niese et al. [2022], which reports placing 50K trees per minute without optimizing functional or aesthetic objectives.Source: §5 · p. 7

6 Conclusion

• GreenScapes combines a constrained image-edit generative prior with a low-dimensional, simulation-driven CMA-ES refinement loop.The generative stage produces stylistically coherent top-down layouts; refinement adjusts scales continuously and substitutes assets discretely within an attribute-space GMM.Source: §6 · p. 7
• The pipeline shows prompt-conditioned controllability and metric-consistent clustering with reconstructed exemplars of known styles.Source: §6 · p. 7
• The authors report substantial refinement gains beyond the generative prior, suggesting a practical route to combine aesthetic intent with surrogate-based performance criteria in automated park design.Source: §6 · p. 7

• The listed authors are affiliated with Adam Mickiewicz University, Purdue University, KAUST, and Christian-Albrechts-Universität zu Kiel.Source: GreenScapes: Optimized Vegetation… · p. 12
• The affiliations span Poland, the United States, Saudi Arabia, and Germany.Source: GreenScapes: Optimized Vegetation… · p. 12
• The listed keywords include urban vegetation, landscape modeling, vision-language models, CMA-ES, and procedural modeling.Source: GreenScapes: Optimized Vegetation… · p. 12

1 Overview

• The supplement outlines conversion of an OSM-derived site representation into a top-down conditioning image, VLM layout prior, metric 3D scene, deterministic repair, and CMA-ES refinement.Source: §1 · p. 12

2.1 Model and API settings

• Reported layout-generation experiments use Gemini 3 Pro image preview; generated samples are stochastic because the API exposes no random seed.The settings describe 1024×1024 generation by default, with downstream stages using fixed seeds and robustness statistics computed over independently sampled layouts.Source: §2.1 · p. 12

2.2 Representative layout-generation prompt

• The representative layout prompt locks non-grass pixels and preserves input geometry, restricting vegetation placement to grass areas.The template also requires distinct plants with visible grass gaps and limits total plant count to approximately 300.Source: §2.2 · p. 13

2.3 Representative arrangement-rule block

• The French formal garden example uses symmetry and geometric patterns, aligns tree rows along paths, and forms shrub outlines from separate plants rather than continuous hedges.Source: §2.3 · p. 13

2.4 Object-annotation prompt

• Instance extraction uses labeled bounding boxes in normalized coordinates, while city-scale examples instead use color-coded dots decoded by thresholding and connected components.Source: §2.4 · p. 13

3 Projection, Feasibility Checks, and Repair

• Each VLM box maps to a 2D instance position and footprint; feasibility checks test the planting mask, lot boundary, obstacles, paths, and vegetation collisions.The footprint is the disc cross-section of a 3D bounding ellipsoid; vertical extents come from class- and asset-specific priors.Source: §3 · p. 14
• Invalid instances are repaired once before optimization by moving their centers to the nearest feasible planting-mask location, then shifting footprints along the signed-distance gradient.Center distance is measured in metric lot coordinates; the second stage addresses remaining obstacle intersections.Source: §3 · p. 14
• Instances still invalid after 30 repair iterations are discarded and excluded from counts, biomass, optimization variables, and validation statistics.During CMA-ES refinement, feasibility is maintained by clamping scale parameters to their bounds.Source: §3 · p. 14

4 CMA-ES Reproducibility Settings

• The active vector has 2M scale variables and, with GMM-based species substitution enabled, 6M whitened attribute variables for M feasible instances; positions are not optimized.All feasible instances in the repaired configuration are included.Source: §4 · p. 14
• The repair algorithm sets a maximum of 30 iterations and discards instances without feasible placement after that limit.Discarded instances are excluded from downstream counts and optimization.Source: §4 · p. 14
• CMA-ES uses a separable variant, re-anchors its mean to the best-so-far solution each generation, and begins with uniform exploration before per-axis adaptation.The initial step size is configured per experiment between 0.15 and 0.4 in normalized parameter space.Source: §4 · p. 14
• Scale multipliers are mapped through a sigmoid, typically bounded to [0.3, 2.0] times default asset size, then clamped by crown-radius and height limits.The same parameterization scheme is used across experiments.Source: §4 · p. 14
• The GMM substitution prior uses 22 species kernels with diagonal covariances and uniform weights; proposals outside every species' Mahalanobis threshold of 3.0 are projected to the closest kernel's threshold ellipsoid.The six-dimensional attribute prior parameters are released in data/species/species.json.Source: §4 · p. 14

5 Objective Terms and Ablation Configurations

• The objective combines sky-view factor, skyline fit, path shading, acoustic exposure, and biomass using weighted squared relative error against targets, with optional scale regularization and feasibility handling.View-dependent terms are averaged over viewpoints; the regularizer penalizes mean squared scale-multiplier deviations from one, with weight 10^-3 unless stated otherwise.Source: §5 · p. 15
• Skyline fit uses flatness, periodicity, or deviation from a target profile; path shading can be target-matched or monotonic, while acoustic exposure measures excess attenuation relative to a free-field reference.The acoustic term is a deterministic optimization surrogate rather than a calibrated wave-acoustic simulation.Source: §5 · pp. 15–16
• Sky-view factor measures the fraction of sky pixels in a circular fisheye domain and is target-matched when a target SVF is specified.The hemispherical view is composited from five cube-map faces using an equidistant projection.Source: §5 · p. 15
• Ablations assign a dominant weight to the selected objective and low or zero weights to the others; reproducibility settings are fixed unless an experiment specifies otherwise.Experiment weights, targets, and step sizes are supplied in released configuration files.Source: Table 2 · p. 15, Table 3 · p. 16
• The biomass term uses asset-specific biomass coefficients and current vegetation scale, with the ablation rewarding higher normalized biomass subject to feasibility constraints.Biomass is computed as occupied canopy volume on the voxel grid scaled by the per-species coefficient.Source: §5 · p. 16

6 Optimization Trace

• Most normalized target errors fall rapidly within approximately the first 10 generations, followed by slower, incremental refinement.The trace is for Joyce Kilmer Park; normalized target error ranges from 0 at target to 1 at worst.Source: §6 · p. 16
• Shade and biomass improvements are occasionally non-monotonic, reflecting trade-offs from discrete asset substitutions and constraint repairs.Source: §6 · p. 16
• By approximately 30 generations, SVF, skyline fit, shade, and biomass reach low residual error, while noise reduction remains higher and harder to satisfy under the same geometric constraints.Source: §6 · p. 16

7 Acoustic-Exposure Details

• Each vegetation asset receives geometric-acoustic descriptors from 20,000 single-bounce rays for each of 16 incident directions.The descriptors are random-scattering fraction, normalized angular spread, and mean outgoing direction.Source: §7 · p. 16
• The scene is voxelized at 1 m resolution, and sound-energy density evolves through material-dependent anisotropic diffusion and decay, with a point-source term.Vegetation descriptors determine diffusion magnitude and direction and increase decay with random-scattering fraction; buildings use high decay, while air uses fixed background parameters.Source: §7 · p. 16
• Most normalized target errors drop rapidly in the first ∼10 generations, while noise reduction improves more gradually and remains a harder objective under the geometric constraints.The trace shows slower refinement and occasional trade-offs later in the run.Source: Figure 1 · p. 17
• With the released species parameters, vegetation base diffusion saturates at the air-derived cap, so species differ acoustically through anisotropy direction and decay.The implementation specifies fixed material mappings and computes reported attenuation as the mean across receivers of the time-integrated energy ratio.Source: §7 · pp. 16–17
• The objective averages excess attenuation at eight path-network viewpoints against a free-field reference with matching simulation settings.This produces a deterministic acoustic-exposure term across optimization iterations.Source: §7 · p. 17
• 5–10 dB of excess attenuation across receivers in a 12 m tree-belt test is qualitatively consistent with the reported 3–8 dB for similar-width belts.The authors say this sanity test checks order of magnitude, not calibration.Source: §7 · p. 17

8 Notes on Reported Counts and Statistics

• Instance counts are computed after annotation, reprojection, feasibility checking, and repair, excluding instances discarded after failed repair.The convention applies to large-scale examples, biomass calculations, optimizer variables, and validation statistics, so proposed but invalid objects are not counted.Source: §8 · p. 17
• Across 147 experiments, 48 VLM-derived initial layouts had at least one invalid placement, and every optimized scene underwent a feasible configuration, scale, or asset/species change.Invalid instances were repaired when possible and otherwise discarded; this supports treating the VLM as a semantic prior rather than the final layout generator.Source: §8 · p. 17

9 Code and Data Release

• The release includes the optimization engine, VLM pipeline, acoustic surrogate, prompts, configurations, species database, analysis tools, site data, and automated tests.Commercial vegetation assets cannot be redistributed; the release provides bounding-volume-matched placeholders, so quantitative results reproduce while renderings appear schematic.Source: §9 · pp. 17–18
• Because the image-generation API exposes no random seed and VLM APIs may change, the release records the model string, prompts, image resolution, and post-processing rules.The reported VLM outputs are stochastic.Source: §9 · p. 18