The “Winner” logic for organic density gradients: Random < Probability, or what I call “Density Bleeding.”

I’ve tested two methods using Grasshopper: Deterministic Culling & Stochastic Sampling.

SDS (Stochastic Density Sampling) Methodology: “Should we shake the sequence, or should we question the existence?”

While both work — and counts are controlled well in the former — the logic of the latter is so beautiful. It’s like designing the “Potential” in a field of survival.

By applying Random < Probability, a scalar field is transformed into an existence filter. It’s a shift toward Field Ontology — sampling the latent potential of space rather than just commanding points. This is how we design “Bleeding” in the age of Generative AI.

Here is the story.

Designing Bleeding — Gradation in Density Random < Probability

In the realm of computational design, point-driven geometry sits at the very core of our work. Whether we are dealing with Voronoi diagrams or circle packing — generative processes that have become staples in our daily workflow — the outcome depends entirely on how points are distributed. This makes the control of point distribution, and specifically point density, the single most critical factor in determining the quality of a design.

The Problem: The Limitation of Grid-Based Density

In architectural and quantitative terms, density is typically controlled by “Count per Area” or “Cells.” However, this approach is inherently dependent on grid systems, which imposes geometric constraints. Specifically, expressing the organic “bleeding” of densities found in nature is remarkably difficult with traditional grid-based addition.

Objective & Research Question

The goal is to establish a grid-independent density control method that manifests a natural “bleeding” gradient.

RQ: Which sampling method is most effective for manifesting natural “bleeding” in a point cloud?

Methodology: SDS (Stochastic Density Sampling)

The SDS (Stochastic Density Sampling) methodology presented here implements the core principles of Stochastic Sampling— the mathematical foundation of modern Generative AI — into architectural density design.

While traditional methods (Count per Area) take an “additive” approach by placing points onto a grid, SDS proposes a paradigm shift: generating a broad “Field” of potential points first, and then performing Subtraction (Culling) by manipulating spatial information.

This study compares two culling approaches — Deterministic and Probabilistic— to evaluate how they manifest the atmospheric quality of “bleeding.” A curve is established as an attractor, where the distance from it defines the transformation parameters of the “Edges.” Density is set high near the attractor and decays as distance increases, with the probability density controlled via a Graph Mapper.

1. Subset Culling (Deterministic Approach)

This method sorts points by distance and extracts the ‘Top N’ subset. It’s the classic logic also introduced by David Rutten — reliable and robust, but I wanted to push further. By experimenting with various noise patterns, I attempted to move beyond this deterministic approach to achieve a true ‘bleeding’ effect.

2. Stochastic Culling (Probabilistic Approach)

An emergent control method applying the core principles of Generative AI. Here, distance is converted into “Existence Potential.” The survival of each individual point is determined by the following autonomous rule:

Random < Probability

By entrusting the materialization of points to an uncertain field, we attempt to “develop” (manifest) true bleeding.

Results

1. Sort + Subset Culling (Deterministic Approach)

1.1 Sharp Culling (Geometric Distance)

Pure sorting by shortest distance. The boundary is extremely sharp, resulting in a geometrically clear and rigid density.

1.2 Shuffle Culling (Sequence Jitter)

The list order is jittered, and points are culled by random indices. This results in uniform “sandstorm-like” noise across the entire field.

1.3 Random Culling (Noisy Distance)

Random noise is added to the distance values before sorting. While this creates some roughness at the edges, the intensity and location of the effect cannot be precisely specified.

1.4 Weighted Random Culling (Weighted Priority)

Weighting is applied via a Graph Mapper to intentionally increase the survival priority of specific regions. While this produces the result closest to “bleeding” within the Subset method, it suffers from a critical flaw: the density gradient is not proportional to changes in the population size. When the population exceeds a certain density, the intended bleeding effect vanishes, and the edges sharpen.

2. Stochastic Culling (Probabilistic Approach)

2.1 Basic Stochastic Culling

Logic: R < P (P: Distance-derived existence probability, R: Uniformly distributed random number).
Result: The decisive difference from the Subset method is that the boundary is so soft it cannot be mathematically defined by a single threshold. Due to the nature of the Uniform Distribution used by the Grasshopper Random component, survival probability decays linearly and uniformly relative to distance.
Data Observation: The points “diminish” naturally according to statistical expectation — reducing a population of 3,000 to 1,811, or 1,000 to 588. However, in this basic setup, the rate of decay depends solely on the distance function and cannot be independently adjusted.

2.2 Weighted Stochastic Culling

Logic: R < f(P) (f: Non-linear distortion via Graph Mapper).
Result: This manifests the most Generative and lifelike density change, resembling a cloud or mist. As the population is culled (e.g., from 4,000 to 848, or 400 to 74), the designer’s intent (weighting) permeates the “stochastic fluctuation,” resulting in an extremely refined density gradient.
Technical Note: For experimental consistency with 1.4, an “Inverted Logic” (P < R) was applied as a shadow logic for comparison.

Discussion

Transformation of “Edges” and the Limit of Determinism

While Subset Culling (Method 1) is practical for manual count control, its reliance on “Top N” extraction creates an unavoidable threshold. No matter how much jitter is added, a hard geometric boundary eventually emerges. Furthermore, it faces a contradiction in scalability: as the resolution (population density) increases, the physical distance between ranked points shrinks, causing the visual “bleeding” to disappear.

The Superiority of Stochastic Culling

In contrast, Stochastic Culling (Method 2) maintains a consistent “bleeding” quality regardless of the input population size. It reproduces an “ink-soaking-into-paper” effect because the sampling logic is tied directly to spatial coordinates (x, y, z)— the Potential Field — rather than list indices. The outstanding stability of this method allows designers to fine-tune density while preserving the atmospheric tendency of the gradient.

Designing the Potential Field: Autonomous Rules

The equation Random < Probability represents a bottom-up “Autonomous Rule” where each dot determines its own fate. When generating organic fluctuations, the designer’s role is not to manage specific “counts” but to design the “Tendency” (Field). By manipulating probability rather than points, we manifest a reality that feels more Tangible and grounded in the logic of nature.

Conclusion: Design the Rule, Not the Form

The SDS Methodology shifts our focus from drawing “Samples” (forms) to designing “Parameters” (rules). We no longer need to command the exact position of every point. Instead, we design the potential for form to emerge and sample the “truth” of that specific moment from an infinite universe of possibilities (Population).

Future Work

The next challenge is integrating the quantitative “Total Count Control” of the Subset method with the probabilistic approach of Type 2. The goal is to build a system that satisfies both the aesthetic beauty of the potential field and the quantitative requirements of architectural programming.

The Grasshopper definition for this project is available in the following GitHub repository

GitHub - cfuyuki/Stochastic-Density-Sampling: Design Bleeding "Nijimi/にじみ" - Implementing Probabilistic Sampling for Architectural Density Design