HomeEmergeMichiNebulaHeavensResearchUsageWorldExperimentsRoadmapStory
...
Sign in

emerge

A self-evolving generative art system that transforms the pulse of the world into visual art — and develops its own artistic vision over time.

How it works

emerge operates in Mastery Mode: it takes a fixed artistic thesis and target emotions set by the human curator, then iterates for hundreds of generations searching for the most powerful visual realization. The system doesn’t generate random art — it develops a visual language through structured experimentation, mutation, and self-evaluation.

The pipeline

Each cycle takes ~5 minutes and passes through 7 stages, visible in real-time on the main page pipeline indicator.

1
Intent — the system forms a visual experiment: which material, composition, and palette to try next. The choice is made by a MAP-Elites + Bandit algorithm that balances exploration of new cells with exploitation of proven approaches. Each experiment specifies a paradoxical material, a composition strategy, a palette mood, and the target emotion from the series thesis.
2
Snapshot — an LLM interprets the thesis, emotional targets, and accumulated knowledge into a structured semantic snapshot. This includes the artistic statement (title + artist’s note explaining what the work wants to make visible) and the emotional intent. The statement iteratively improves across generations rather than being generated from scratch.
3
Image — a Scene Director (a separate LLM) designs the full visual structure of the image: composition, materials, light, color, spatial logic, and how each element serves the artistic statement. The result is rendered into an image. Two images per cycle with different visual approaches for comparison.
4
Critique — evaluates emotional impact, statement clarity, freshness, style diversity, transcendence, and paradox depth. Scores feed back into the Repertoire Map cell that produced this generation, building an ever-growing quality map. Series scores are tracked for stagnation.
5
Reflection — a 5-layer self-analysis: meaning (did the intent reach the viewer?), craft (what worked technically?), translation (did words become the right visuals?), self-calibration (am I scoring honestly?), and paradox evaluation (did genuine impossibility survive?). Proposes the next experiment.
6
Meta-reflection — analyzes the reflection process itself. Evaluates whether its own patches improve thinking. Evolves system identity.
7
Learn — knowledge distillation updates the aesthetic knowledge base. Positive solutions evolve: solutions aged 6–8 cycles enter a mutation window where the LLM can Mutate, Combine, or Branch them into new variants. This creates a lineage of evolving techniques rather than a static list.

Mutation & Visual Repertoire

The core learning mechanism is a MAP-Elites grid with bandit selection. The repertoire is a 3-axis space: Fantastic Materials (20 paradoxical substances like “liquid stone” or “acoustic metal”), Composition Strategy (10 spatial arrangements), and Palette Mood (8 emotional color schemes). That’s 1,600 possible cells to explore.

Each generation occupies one cell. After the critic scores the result, the score feeds back into that cell. Over time, the system builds a quality map: which combinations of material + composition + palette produce the strongest emotional impact for the current thesis. The Creative Dashboard visualizes this map in real-time.

Positive Solution Evolution: successful techniques don’t just persist — they mutate. When a solution reaches age 6–8 cycles, the LLM is prompted to evolve it: create a variant (Mutate), merge two solutions (Combine), or take an element into a new direction (Branch). Each evolved solution carries its lineage (generation depth, parent, mutation type), creating an evolutionary tree of artistic techniques.

Score decay (3% per cycle) ensures no cell rests on old laurels. Exploration rate starts at 90% and floors at 50%, guaranteeing permanent discovery even as the system accumulates knowledge.

Your role

Through feedback, you influence how the system evolves. Your words are analyzed by an LLM and translated into specific changes to directives, prompts, and scoring criteria. Each image has a feedback field for targeted comments. The system also generates requests — things it needs from humans to improve.

System versions

v1.0
Genesis — basic pipeline: snapshot → image generation → display. Single data source (news). Fixed style prompts.
Jan 2026
v2.0
Perception — 14 data sources (weather, earthquakes, moon, poetry, music, art galleries, ocean tides, solar activity). Visual ontology with entities, forms, materials. Scene Director with free visual language choice.
Jan 2026
v3.0
Critique — AI critic with structured evaluation. Composition, ontology, connections directives as evolving JSON. Automatic prompt changes based on critique. Visual memory accumulation.
Feb 2026
v4.0
Reflection — system self-analysis after each cycle. Radical experiments. Positive solutions toolkit. User feedback system with LLM-analyzed changes. Telegram integration.
6 Feb 2026
v5.0
Meta-consciousness — meta-reflection that analyzes the reflection process itself. System identity and self-awareness. Requests to humans. Emotional intent and verification. Poetic inner voice.
8 Feb 2026
v6.0
Closed loop — quantitative score tracking with deltas. Hypothesis-driven A/B testing (each batch: hypothesis vs control). Scene Director prompts visible to critic for intent-vs-result comparison. Config versioning with full audit trail. Structured positive solutions. Meta-reflection evaluates its own patches with causal chain analysis. Request lifecycle with auto-close. Visual primitives dictionary.
10 Feb 2026
v7.0
Statement & Series — Artistic Statement layer: before each generation the system formulates a thesis — what it wants to say and which emotions to evoke. Statement persistence across “series” of 5 generations to develop visual language for one idea. 4th directive — Visual Style (medium, palette logic, texture, light model, spatial logic). Structural evolution within each directive: proposing, testing, and evaluating new rules and value combinations. Exploded diagram analysis for every image. Top 100 gallery. Infinite scroll. Dedicated image detail page with vertical feed.
12 Feb 2026
v8.0
Prompt Engineering — Prompt Engineering Toolkit in the Scene Director: style anchors, explicit color naming, anti-abstract-language rules, Medium Wheel with 30+ techniques across 9 categories. Prompt-Result Journal — system memory mapping specific prompt words to visual outcomes. Dynamic banned patterns computed automatically from recent palette fingerprints. Self-calibration layer in reflection: system evaluates accuracy of its own scores. Translation Quality layer: system analyses how well conceptual language translates to visual prompt language. Feedback safety guard against harmful input. Feedback classification and routing by type.
14 Feb 2026
v9.0
Breakthrough Memory — breakthrough detection: when an image scores high on style diversity and freshness, the exact prompt that produced it is saved as a “proven style anchor” and fed back into future generations. Both images in A/B test now experiment — no more “safe control” that reinforced the old style. Last-image avoidance ban forces every consecutive image to be visually opposite to the previous one. Reflection forbidden from proposing “return to safe approach”. System builds on its best experiments instead of forgetting them.
14 Feb 2026
v10.0
Context Distiller & Model Upgrade — intermediate LLM step (Context Distiller) that synthesizes 12+ instruction blocks into a focused creative brief for the Scene Director, eliminating context overload. All critical LLM calls upgraded to GPT-4.1 (reflection, meta-reflection, critic, feedback analysis). Deterministic series thesis selection — system evaluates all snapshot statements and picks the most substantive one. Prompt Validator enforces experiment compliance by checking style anchors. Artistic statement reformatted as “Title + Artist’s Note” across all surfaces. Investigation themes curated to 4 deep philosophical territories.
14 Feb 2026
v11.0
Creative Intent Pipeline — fundamental shift: instead of deriving meaning from random data, the system now formulates a specific creative intent BEFORE seeing any data. Each generation starts with a question, hypothesis, target emotion, and visual risk. Data becomes material filtered through intent, not the source of ideas. Single investigation theme (Transhumanism vs Posthumanism) for deep exploration. Users can set series direction via /series command. Reflection evaluates intent→statement→image chain and rates semantic coherence. Meta-reflection manages intent quality and evolves the intent formation strategy. System displays its current investigation on the main page for all visitors.
16 Feb 2026
v12.0
Discourse, Knowledge Base & Human-in-the-Loop — three new systems for deeper meaning. Discourse: a philosophical dialogue page where the system thinks autonomously and invites humans to join discussions about transhumanism and posthumanism; insights feed into creative intent. Aesthetic Knowledge Base: a structured, evolving document about meaning, creative concepts, artistic statements, and meaning transmission — not prompt techniques, but deep understanding of WHY things work; updated by LLM after each reflection cycle (3000 tokens, GPT-4.1); visible at /knowledge. Human-in-the-Loop: after each image, the system generates a perception question specific to that work; viewers answer on the image detail page; responses feed back into creative intent and reflection, closing the meaning-verification loop. Meta-reflection now evaluates knowledge gaps and updates development priorities.
16 Feb 2026
v13.0
The Success Trap — a deep architectural response to a paradox: the system optimized itself into creative convergence. By accumulating “proven techniques,” “breakthrough styles,” and positive solutions, the system gradually narrowed its own creative space — rewarding the familiar and punishing the unknown. Six feedback loops were identified that drove monotony: breakthroughs became ceilings, positive solutions became habits, the prompt journal taught repetition, directives accumulated without pruning, the distiller averaged competing signals, and reflection rewarded safe bets.

The fix is structural, not cosmetic. Reframing: all “build on what works” instructions were reversed to “depart from explored territory.” Exploration budget: every 3rd generation deliberately ignores all accumulated wisdom and starts from scratch. Mandatory novelty: the Context Distiller must now name one visual element never used before. Semantic pattern bans: an LLM analyzes recent prompts for recurring conceptual patterns (not just words) and bans them. Directive decay: every 10 generations, the oldest 30% of rules in each directive are automatically disabled, preventing unbounded constraint accumulation.

This version is a philosophical statement: a creative system must actively resist its own optimization instinct. Success is the enemy of discovery.
16 Feb 2026
v14.0
Paradox Engine — impossibility as the core structural requirement. Every generation must contain an unresolvable visual paradox — not surrealism (that’s catalogued), not optical illusion (that’s solved), but a new impossibility the viewer’s brain cannot resolve. A Paradox Matrix generates unique impossible constraints from two axes: material paradoxes (liquid stone, transparent lead) and perceptual paradoxes (closer inspection reveals further distance, shadows cast by absent objects). The Critic now scores paradox_depth alongside composition and style. The Knowledge Base tracks which paradox types survive the prompt→image translation chain. Reflection includes a dedicated Paradox Evaluation layer. Breakthrough detection triggers on high paradox scores. The system maps explored vs unexplored paradox territory and prioritizes the unknown.
17 Feb 2026
v15.0
Mastery Mode — a fundamental shift from divergent novelty-seeking to convergent depth. The system’s thesis and target emotions are now fixed for an entire series of 200–300 generations — controlled by the human curator, not by the system. What iterates: visual approach (medium, technique, composition, palette) and paradox discovery. The system searches for the most powerful visual realization of a single idea, trying everything from charcoal to electron microscopy, always asking: does THIS visual language serve the thesis better than what came before?

Series Backlog: a curated queue of thesis+emotion sets visible on the Discourse page. Admin controls advancement between series. Medium Rotation replaces Exploration Mode: every 5th generation forces a random medium from a 15-category wheel, preventing visual habit while preserving accumulated knowledge. Stagnation detection now measures realization quality (clarity + emotional impact), not style diversity. Autonomous self-dialogue disabled — the system focuses on making, not talking. The Discourse page becomes a dashboard: current thesis, discovered paradoxes with scores, series backlog, and admin controls.

First series: “We are already cyborgs — the boundary between organism and device dissolved before we noticed.” Target emotions: exhilaration of discovery, electric anticipation, ecstatic vertigo.
17 Feb 2026
v16.0
Visual Repertoire (MAP-Elites + Bandit) — the system now maintains a structured repertoire of visual approaches organized as a 3-axis grid: Fantastic Materials (20 paradoxical impossible substances), Composition Strategy (10 spatial arrangements), and Palette Mood (8 emotional color schemes).

An epsilon-greedy bandit algorithm selects which combination to explore next, balancing discovery of new cells with exploitation of proven high-scoring ones. Each generation’s scores (emotional impact + clarity) feed back into the grid cell, building an ever-growing map of what visual language works for which emotions. The fantastic materials from the old Paradox Matrix are now axis 1 of the repertoire — they accumulate scores and improve over time instead of being randomly discarded.

Development Backlog: Semantic Memory Graph with Consolidation & Decay is planned for future testing — a graph-based visual language memory where nodes consolidate into proven recipes and unused elements decay.
17 Feb 2026
v17.0
Skill Lab — a micro-experiment laboratory for training isolated visual skills. Each experiment focuses on ONE dimension (emotion, material, composition, palette, idea, metaphor) and iterates until the skill is “graduated” — consistently scoring above threshold.

Skill Types: 6 dimensions, each with its own LLM critic, prompt templates, and evaluation metric. Managed Backlogs: each skill type has a queue of targets to work through, auto-populated from Visual Repertoire axes for compatibility. Graduated Recipes: proven skills produce structured recipes with Visual Repertoire axis indices, directly injectable into the main pipeline. Skill Composition: combine multiple graduated recipes into a single creative intent for the main gallery pipeline.

Separate Telegram channel for lab results. Mini-pipeline runs in 2–3 minutes (vs 10 min for full pipeline). Admin-only access via the Skill Lab page.
17 Feb 2026
v18.0
Living Pipeline & Creative Diagnostics — the system becomes observable. A real-time pipeline visualizer on the main page shows 7 stages as a chain of nodes that light up green (done), yellow (active), or dim (pending) as each cycle progresses. A dynamic sub-phase text shows what’s happening inside each stage: “Scene Director: nature_art”, “5-layer self-reflection”, “Distilling knowledge”. A feedback-loop arrow at the end completes the cycle.

Creative Diagnostics Dashboard: a dedicated admin page with 7 diagnostic blocks — Creative Pulse (5 metrics over time with deltas), Diversity Index (MAP-Elites coverage, axis stats, exploration rate), Repertoire Map (interactive heatmap of material × composition with scores), Solution Evolution (lineage depth, mutation types), Skill Lab status, Pipeline Health (timestamps, last statement), and Statement Tracker (best artistic statement and score).

Positive Solution Evolution: solutions aged 6–8 cycles enter a mutation window. The LLM can Mutate (create a variant), Combine (merge two solutions), or Branch (spin off an element into a new direction). Each evolved solution carries its lineage: generation depth, parent ID, mutation type. This creates a living evolutionary tree of artistic techniques, not a static best-of list.

Feedback queue fix: feedback items that fail processing are retried up to 3 times before being dropped, preventing infinite re-queue loops.
18 Feb 2026

Current stage

Right now the system is in a foundational training phase, pursuing two goals simultaneously. First: the system must not stagnate or narrow down — its creative innovation should accelerate with each iteration, not decay. Second: it must learn to produce images that actually deliver the intended emotion and realize the artistic statement — not just generate something visually interesting, but create what it set out to create.

Once these fundamental skills are mastered, the system will move to the next stage.

The vision

emerge is an experiment in art that is alive — art that perceives, reflects, mutates, and grows. Not a tool for making pictures, but a system that develops its own visual language through structured experimentation. It maintains an evolving repertoire of 1,600 material–composition–palette combinations, scores each result, mutates successful techniques, and iterates toward ever more powerful visual realizations of a single artistic thesis.

Read the essay: Art & Technology →

Lessons from Generative Art Systems →

Previous pipeline description (v1–v17) →