Neural Morphogenesis

Form from one learned local rule. Every living site applies the same tiny network to its neighbourhood, so a seed can grow a patterned body that persists—and sometimes regenerates—without a runtime blueprint. The spectacle is not a single drawing pass, but a rule that can carry identity through time and injury.

Neural Cellular Automata (NCA) extend classical cellular automata by replacing a hand-written lookup table with a learned local update. In the Distill Growing Neural Cellular Automata lineage, each cell state is a vector. Perception gathers local information; a small multilayer perceptron proposes a residual change; stochastic updates and an alive mask keep growth from becoming a rigid global iteration.

The 0.8.0 pack follows that architecture closely:

• State: twelve channels per cell.
• Perception: for each channel, the cell value plus Sobel gradients in (x) and (y) — three features × twelve channels = a 36-dimensional perception vector.
• Update: an MLP maps (36\rightarrow64\rightarrow12) and adds the result residually to the previous state.
• Fire rate: each cell updates with probability fireRate; otherwise it largely retains state (with vitality-related nudges under the alive mask).
• Alive mask: typically keyed to an alpha-like channel and its neighbourhood, so “dead” cells do not keep inventing tissue from empty space without seed or neighbour support.

Training (offline) teaches different behavioural families. The shipped pack does not reload Mordvintsev et al.’s published Distill checkpoints. Weights are project-trained int8 modules pinned to revision nca-2026-07-28.4. Public copy must therefore say Distill-style rather than identical to the Distill demo.

Vitality is not metabolism

The simple control Vitality stimulates alive-masked hidden channels so the field does not freeze into a lifeless still image. It is a numerical life-support for the automaton, not a model of nutrient flow, ATP, or developmental signalling. Raising vitality can re-energise pattern dynamics; it does not feed a body.

Damage, seeds and brushes are experiments on the rule

damageRate randomly clears disks. Brush targets Cut, Transplant (graft), Poison (perturb hidden channels) and Add seed let visitors ask whether the local rule rebuilds, fragments or invents. These interventions test distributed memory in the field, not surgical biology.

Cellular automata (von Neumann, Ulam; later Conway’s Game of Life) showed that local discrete rules can generate complex global structure. Neural Cellular Automata replace fixed tables with differentiable local networks. The Distill article Growing Neural Cellular Automata (Mordvintsev, Randazzo, Niklasson and Greil, 2020) made growing, persistent and regenerating NCA vivid as interactive artefacts and as a research programme in self-organisation and regeneration.

That lineage is conceptual and architectural. ecospheres.ai’s Neural Morphogenesis sphere implements the same family of update rules with its own trained weights, packing format and public surface. It is not a port of one Distill checkpoint file.

Adjacent traditions—reaction–diffusion morphogenesis, Lenia’s continuous organisms, Cellular Potts tissues—share the atlas’s broader interest in form from local interaction, but they are different mathematical objects. NCA is distinctive because the local rule is a learned function rather than a hand-specified PDE or energy.

Aspect	Current implementation
Entities / field	Square grid of cell state vectors (twelve channels).
Perception	Sobel×3 per channel → 36-D local feature vector.
Update	Residual MLP (36\rightarrow64\rightarrow12); stochastic fire rate.
Alive mask	Neighbourhood-aware mask (alpha channel lineage) gating empty growth.
Weights	Project-trained int8 modules; revision nca-2026-07-28.4 (not Distill publish checkpoints).
Models (modelId)	growing-v1, persistent-v1, regenerating-v1, texture-v1, random-v0.
Seeds	Centre seeds on reset; count controlled by seedCount.
Ongoing damage	Optional random disk clears via damageRate.
Vitality	Alive-masked hidden-channel stimulus; not metabolism.
Step policy	Fixed ~1/60 s policy with limited substeps; optional worker with main-thread fallback.
Rendering	Palette mapping of visible channels; advanced hidden/contrast/trail options exist off simple surface.
Public simple controls	Model, Fire rate, Seed count, Damage rate, Vitality, Palette.
Public simple tool	Brush only, with targets Cut / Transplant / Poison / Add seed.
Hero framing	Host observer camera enabled (Reset view); presentation-only.

The most truthful one-line description is: a Distill-style neural CA whose project-trained local rule grows and maintains patterned bodies from seeds.

Growth from almost nothing

In Growing Seed, a minimal seed expands into a coherent body. Watch whether structure appears gradually from the seed outward rather than as a stamped sprite.

Persistence versus continual remaking

Persistent Body asks whether the rule can hold a form. Look for a stable silhouette that still micro-updates, not a frozen bitmap.

Regeneration after injury

The Regenerator and Distributed Memory, especially with damage or Cut, test whether cleared regions refill from remaining tissue. Regeneration here means the local rule reconstructs pattern, not that cells “heal” biologically.

Twin seeds and competition

Twin Embryos places two seeds. Observe merger, coexistence or interference—evidence that the rule is spatial and local, not a single global template blit.

Alien texture regimes

Alien Tissue (texture-v1) explores non-figurative grown textures. Treat unfamiliar silhouettes as alternate attractors of the same architecture, not as species taxonomy.

Fire rate as a temporal microscope

Lower fire rate slows the stochastic update. Pattern becomes easier to read; growth and repair take longer. Higher fire rate densifies updates and can make motion look more fluid or unstable depending on the model.

First 30 seconds

1. Load the showcase Growing Seed.
2. Identify the seed and watch local growth without touching controls.
3. Switch Model to Persistent Body’s weights (or load that preset) and compare stillness versus ongoing micro-motion.
4. Load The Regenerator. Use Brush → Cut, then watch whether the gap refills.
5. Try Add seed once away from the body; ask whether a second growth competes or merges.
6. Nudge Vitality only after you have a baseline—treat it as stimulus, not “feeding.”

Three experiments

Experiment	Question	Do	Watch for	Why it matters
Rule, not stamp	Is the body a sprite or a process?	Growing Seed → Cut a notch → wait.	Local refill or collapse from the cut edge.	Shows the update is generative and ongoing.
Damage as pressure	How much injury can the rule absorb?	Regenerator → raise Damage rate gradually.	Persistent scars, collapse or continual repair.	Separates robust regeneration from fragile growth.
Two seeds	Does one global shape win?	Twin Embryos or Seed count 2 → observe.	Merger, twins or interference fringes.	Demonstrates spatially distributed causation.

Parameter	What it really controls	Increase it	Decrease it	What to watch
Model (modelId)	Which trained weight set runs.	Categorical.	Categorical.	Entire behavioural family: grow / persist / regenerate / texture.
Fire rate	Probability a cell applies the MLP update.	Denser temporal updates.	Slower, more readable evolution.	Growth speed and flicker of micro-change.
Seed count	Number of centre seeds at reset.	More initial growth centres.	Single-origin growth.	Merger versus multiple bodies.
Damage rate	Probability of random disk clears per step.	Chronic injury.	Intact development.	Whether the rule outruns damage.
Vitality	Stimulus into alive-masked hidden channels.	Stronger ongoing activation.	Greater risk of freeze/stillness.	Motion without interpreting metabolism.
Palette	Colour mapping of visible channels.	Categorical aesthetics.	Categorical.	Appearance only; not a new rule.
Resolution (advanced / plate)	Grid size.	Finer bodies, higher cost.	Coarser, faster.	Numerical ontology, not cell biology.
Damage radius (advanced)	Size of random damage disks.	Larger wounds.	Pinprick injuries.	Scale of regeneration challenge.
Show hidden (advanced, visual)	Expose hidden channels.	More channels visible.	RGB-like reading.	Instrument for seeing state, not a new physics.

The best first pair is Model and Fire rate: one selects the learned behaviour, the other sets how quickly that behaviour is sampled.

Growing Seed

Showcase on desktop and mobile. The clearest introduction to seed → body growth under growing-v1.

Persistent Body

Switches to persistent-v1 with slightly higher vitality. Ask whether form holds without continual dramatic remaking.

The Regenerator

regenerating-v1 with elevated fire rate and a low damage rate. Best first regeneration classroom.

Twin Embryos

Two seeds under the growing family. Use it to teach spatial locality and interaction between growth centres.

Distributed Memory

Regenerating weights under stronger damage pressure. The title is editorial: the field stores pattern in distributed state, not in a separate memory bank object.

Alien Tissue

texture-v1 with aurora palette and lower fire rate. A reminder that the architecture is not limited to cute figurines.

Recommended learning order: Growing Seed → Persistent Body → The Regenerator → Twin Embryos → Distributed Memory → Alien Tissue.

• Brush is the only tool on the ecospheres.ai simple surface.
• Cut clears tissue in the brush.
• Transplant (graft) copies a local patch—useful for asking whether displaced tissue reintegrates.
• Poison perturbs hidden channels without necessarily erasing the visible body.
• Add seed deposits seed material to start or challenge growth.

There are no field actions on the simple surface. Spawn/erase exist in the wider pack tooling but are not exposed on the current ecospheres simple configuration for this engine.

This is not biological morphogenesis. There are no genomes, morphogen PDEs, cell lineages or tissues in the wet-lab sense.

Weights are Distill-style, not bit-identical Distill publish checkpoints. Do not claim reproducibility against a specific Distill demo file.

Vitality is not metabolism, hunger or health.

Palette and trails do not change the learned rule.

Regeneration is pattern reconstruction by a local update, not wound healing with inflammation or stem-cell biology.

random-v0 and texture models are not “species.” They are alternate weight sets.

The defensible conclusion is:

A single learned local rule, applied everywhere with perception and an alive mask, can grow, maintain and repair patterned form from seeds — a computational morphogenesis.

Neural Morphogenesis adds learned local rules as a generative principle. Turing Patterns and Excitable Waves show hand-specified continuum interactions; Lenia shows continuous automata with designed growth maps; this sphere shows that a tiny neural update can be trained into growth, persistence and repair. It is the atlas’s clearest invitation to ask how much global form can be delegated to identical local computation.

Foundational and primary sources

1. Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson and Michael Levin / Distill contributors as credited, “Growing Neural Cellular Automata,” Distill (2020). Available at https://distill.pub/2020/growing-ca/. Conceptual and architectural lineage for perception, residual MLP updates, fire rate and alive masking.
2. Classical cellular-automata lineage (von Neumann/Ulam; Conway) as background for local discrete worlds — cited as historical context, not as the NCA training method.