synthetic.pics
GalleryCollectionsCustom GalleryGenerative ArtMethodLicensingFAQ
GalleryCollectionsMethod
Explore
  1. Home
  2. /Method

Human-guided. Machine-generated.

How We Make Generative Art

Synthetic.pics is not a prompt box. We build an art-direction system that decides how the engine explores form, composition, color, surface, and space — then generates and compares possibilities within that territory before rendering the selected directions into artwork.

We build the system. The system explores the possibilities. We preserve the work that earns a place in the gallery.

The complete process

Every artwork passes through the same seven stages, in order. Each stage narrows the field; nothing skips ahead.

  1. Direction→
  2. Explore→
  3. Evaluate→
  4. Art-direct→
  5. Generate→
  6. Validate→
  7. Archive

synthetic.pics

AI art gallery · Human-guided. Machine-generated.

A curated archive of original AI-generated artwork — AI art images, synthetic pictures and AI paintings. New AI artworks added regularly.

Explore

  • AI Art Gallery
  • AI Art
  • AI Art Collections
  • AI-Generated Art
  • Abstract AI Art
  • Minimalist AI Art
  • Digital AI Art

Create

  • Custom Galleries

About

  • Method
  • Licensing
  • FAQ
  • Privacy
  • Terms

Technical

  • Agent Access
synthetic.pics — a curated AI art gallery of AI-generated artworks.·Designed and built by Pantaleone.net
  1. 01

    Direction. Set the visual problem the work should solve.

  2. 02

    Explore. Build candidate art directions from the creative grammar.

  3. 03

    Evaluate. Keep what is coherent, prefer what is new.

  4. 04

    Art-direct. Translate the winning direction into a working prompt.

  5. 05

    Generate. Render the artwork with the image model.

  6. 06

    Validate. Check the result against the archive before it ships.

  7. 07

    Archive. Preserve the work and let it inform what comes next.

01

Direction: define the creative problem

The engine begins with an artistic strategy, not a list of objects to combine. A direction states what the composition is trying to do: set two geometries against each other, distribute visual mass unevenly, interrupt a repeated pattern, or place a single form in open space.

Strategies are sampled with adaptive weighting, so overused ones recede and quiet ones get their turn. The relationship matters more than the objects — say, a rigid geometric structure cut by one irregular translucent form.

02

Explore: build an art direction

Each candidate direction is assembled from the engine’s creative dimensions. These interact — a different material alone does not make an otherwise identical composition a different artwork — and the frame’s arrangement carries weight in every decision.

Strategy
the central visual proposition.
Forms
the principal geometry and its secondary elements.
Composition
placement, scale, rhythm, balance, cropping, and negative space.
Color
palette, value, saturation, and where color appears.
Surface
painterly marks, fine lines, roughness, translucency, layering.
Material and light
physical characteristics and how they interact.
Optical behavior
interference, perceptual tension, or ambiguous space — used sparingly.
Deviation
an intentional departure that can make a familiar structure surprising.

This vocabulary is an expandable creative grammar, not a fixed list of styles. Related forms share families, so renaming a shape doesn’t fool the engine into seeing something new.

03

Controlled randomness: explore without losing direction

Every run starts from a master seed. Independent sub-seeds drive each creative dimension separately, so a palette change doesn’t shift the composition, and each decision stays traceable. The same seed under the same configuration and archive state produces the same direction.

Two directions can share a palette and still differ in structure — a cluster pressed to an edge versus thin lines across the full canvas. And two images in different colors still repeat when they share silhouette, focal position, and arrangement. The engine scores both cases accordingly.

04

Evaluate: preserve coherence and seek novelty

This is the center of the method. The engine develops sixteen candidate directions per artwork and judges them on two competing objectives.

Coherence asks whether the decisions work together. Material, light, color, form, and composition must support one visual relationship — a quiet field cannot carry a high-contrast treatment. Each candidate starts at 100 and loses points per violated rule; minimal work must clear a higher bar than experimental work.

Novelty asks whether the direction adds something structurally different from the archive — strategy, form family, composition, surface, spatial relationships first, color weighted lower. A new palette on a repeated composition does not pass.

The winner balances novelty against recent-history penalties among candidates that clear the coherence bar; if none clears it, the most coherent is kept. Related geometries share form families, overused strategies are down-weighted, and removed works are remembered at full weight — because renaming a composition does not make it new.

Contrast — tiled field in gray, sparse arrangement

Contrast

2026-11-20

Contrast on a quiet ground.

05

Art-direct: turn the direction into a usable prompt

The image model needs a coherent description of the intended result, not a dump of attributes. A language model writes the prompt from the winning direction; if that path fails, a deterministic fallback produces it from fixed structures, so a broken prompt never slips past the safeguards.

Every prompt is validated before rendering: a physical visual description of appropriate length, with no people, no text, no quality-tag filler, and no borrowed artist names. Failures are repaired or regenerated within a bounded budget.

Illustrative example — not a production prompt

Intent
interrupt an orderly structure.
Composition
an offset grid on one side of the canvas, leaving an uneven field of open space.
Forms
rigid rectangular divisions crossed by one irregular, translucent form.
Palette
bone white, deep blue, and one restrained accent.
Surface
crisp edges with subtle material variation.
An asymmetrical geometric composition with an interrupted rectangular grid pressed toward the left edge. A single translucent irregular form cuts across the rigid divisions. Bone-white space dominates the right half, while deep blue defines the grid and a small accent marks the point of interruption. Crisp boundaries and restrained surface variation maintain clarity.

The resulting image is not guaranteed to match the example. The prompt establishes direction; the image model renders its interpretation.

06

Generate: render the artwork

Generation is the rendering stage, not the whole method. The production path uses FLUX.2 Klein at 1024 × 1024, driven by the art-directed prompt and the run’s master seed.

Finished renders pass through a contrast grade and are stored as modern compressed images. Watermarks are display-only: the stored bytes stay clean.

07

Validate: protect the quality of the archive

Before an artwork is published, it passes automated safeguards — a floor that keeps the archive honest, not a guarantee of artistic success:

  • Prompt coherence and validation before any pixels are rendered.
  • Title grounding: every content word of the title must appear in the prompt or the direction’s attributes — an ungrounded title blocks publication.
  • Description grounding: the description’s subject must name something actually in the artwork, or no description ships at all.
  • Near-duplicate titles rejected against recent works.
  • Similarity and novelty checks against the archive, including remembered removals.
  • Metadata integrity and deterministic traceability on every record.

A new title or color on a repeated composition is still a repeat, and the gates are written accordingly. Publication is automatic, so the gates run before rendering. Anything that slips through can be removed, and removals teach the engine. Practical questions about the gallery are answered in the FAQ.

08

Archive: make each result part of a larger body of work

The archive is not a folder of outputs. Every published work joins the novelty baseline the next run is judged against.

Titles, descriptions, tags, and stable records make each work discoverable and traceable, connecting the curated collections and individual artwork pages back to the direction that produced them. The archive serves two purposes: it preserves the work, and it helps the engine avoid repeating its own recent decisions. Broader context on the practice lives on the generative art and abstract art pages.

The expanded creative grammar

The engine selects compatible combinations from the dimensions below — restraint in selection is part of the method.

Composition and space

The arrangement of masses, voids, and edges is a primary source of difference between works.

  • Asymmetric balance
  • Interrupted grids
  • Extreme cropping
  • Nested spaces
  • Dense clusters in open fields
  • Isolated focal points
  • Competing centers
  • Expansive negative space

Forms and geometry

Related structural families exist to produce different visual systems, not different object names.

  • Geometric systems
  • Irregular intrusions
  • Organic volumes
  • Fragmented structures
  • Branching filaments
  • Folded planes
  • Layered transparencies

Color and contrast

Value, saturation, distribution, and interaction matter — not only hue. Shared palettes can still differ completely in weight.

  • Near-monochromatic fields
  • Restrained single accents
  • Large color fields
  • Tonal shifts
  • High-contrast structures

Surface and material

Surface behavior is chosen independently of material, so the same material can appear in genuinely different ways.

  • Painterly marks
  • Fine-line systems
  • Layered veils
  • Translucency
  • Granular texture
  • Abrasion and relief

Optical and spatial tension

Controlled perceptual effects — ambiguity, interruption, impossible space — give quiet compositions something to hold onto.

  • Interference patterns
  • Figure-ground ambiguity
  • Interrupted contours
  • Unusual scale relationships
  • Impossible spatial relationships

Three problems, worked through

Methodology is abstract until it meets actual work. Each example states a creative objective, the choices it opens, how coherence holds, and how the direction differs from a conventional composition — beside an archive work that shows the principle.

Example A

A structured geometric field disrupted by one irregular form. The tension is in the relationship between order and disruption, not in added objects.

Directional choices

  • Grid density and how far the grid is pressed toward an edge
  • Focal position of the interruption
  • How much negative space is left open
  • Color contrast between the system and the intruder
  • The scale of the interruption — a nick or a rupture

Coherence. The grid and the intruder must share a material and lighting logic, or the image reads as two pictures stapled together.

Difference. A conventional version centers a neat grid and decorates it. This one commits to the break: the irregular form crosses the divisions and the open space stays uneven.

Bands — parallel bands in gray, layered arrangement

Bands

2026-11-19

Bands on a muted field.

Order, interrupted — from the archive: Shown as an illustration of the principle — the direction described here is illustrative, not this work’s production record.
  • organic-abstraction
  • interruption
  • parallel-bands
  • muted
  • quiet

Example B

A large restrained field carries an uneven concentration of visual mass. The same colors can still differ in weight and organization.

Directional choices

  • Proportions of the color field
  • Edge behavior — hard boundary or slow bleed
  • Local density: where the mass gathers
  • Surface texture and tonal depth
  • Position of the dominant mass within the field

Coherence. With so little in the frame, palette must match the sampled values — a quiet field cannot carry high contrast, and the mass must sit somewhere deliberate.

Difference. A repetitive version centers a soft shape on a wash. A distinct one pushes the mass off-center, lets one edge go granular while the other stays clean, and leaves most of the field empty.

Primary — scattered marks in blue, centered arrangement

Primary

2026-11-19

Stained necklace holding the middle.

Atmosphere and mass — from the archive: Shown as an illustration of the principle — the direction described here is illustrative, not this work’s production record.
  • atmospheric-abstraction

Example C

A precise optical or geometric system whose rhythm breaks near one edge. Small structural changes outweigh color changes.

Directional choices

  • Repetition interval and line density
  • Underlying geometry of the system
  • Where the break lands — edge, corner, or focal point
  • Cropping: what continues past the frame
  • Contrast between the system and its deviation

Coherence. The system must read as a system first, or the break has nothing to break. One deviation governs the image.

Difference. A conventional pattern fills the frame evenly. This one sets the rhythm, then lets a single violation play out near one edge.

Arcs — curved bands in blue, scattered arrangement

Arcs

2026-10-10

Arcs on a dark background.

Pattern under pressure — from the archive: Shown as an illustration of the principle — the direction described here is illustrative, not this work’s production record.
  • geometric-abstraction
  • ordered-field-controlled-irregul
Composition — stacked slabs in ochre, dense arrangement

Composition

2026-11-20

Composition on a dark background.

Gradients — rectangular planes in ochre, balanced arrangement

Gradients

2026-11-20

Gradients on a dark background.

What the current engine does differently

Earlier versions sampled attributes more independently — mode, form, material, palette — and compared candidates mostly by matching names. The current engine adds structure: a visual proposition first, composition as a first-class citizen, related forms in families, deterministic exploration, compatibility rules, and novelty measured across structural dimensions. Older gallery were made under earlier versions; the engine is versioned, and every record carries the version that produced it.

AreaCurrent direction
Creative starting pointA visual proposition rather than a list of objects.
CompositionA major source of diversity, not a secondary attribute.
Form vocabularyRelated structural families, not isolated names.
RandomnessDeterministic exploration with deliberate variation.
CoherenceCompatibility rules that preserve the visual logic.
NoveltyComparison across structural dimensions and the archive.
Prompt generationStructured art direction translated into a concise directive.
EvaluationMeasured diversity and quality rather than attribute counts alone.

The main principle, in plain terms: two images are not meaningfully different just because one uses a different material name or color.

The human and the machine

The human establishes

  • Visual principles and artistic objectives
  • Creative constraints and compatibility rules
  • System behavior — what the engine may explore
  • Validation gates and archive policy

The machine performs

  • Structured exploration of combinations
  • Coherence and novelty evaluation
  • Prompt writing and image rendering
  • Automated validation before publication

Neither role replaces the other. Art direction decides which possibilities are worth exploring; automated selection reduces repetition but cannot establish artistic value.

Limitations, stated plainly

Credibility comes from describing the method honestly:

  • A unique seed does not guarantee a visually unique image.
  • Different prompts can produce similar compositions.
  • A coherence score measures compatibility of decisions, not universal artistic quality.
  • Similarity checks reduce repetition but cannot catch every visual repeat — they work from direction descriptions, not pixels.
  • The image model renders its own interpretation of every prompt.
  • The vocabulary and selection rules need continued evaluation; not every result is suitable for publication.

The system does not understand art the way a human does. It creates the conditions for distinctive work and keeps the gallery from repeating itself — it does not guarantee originality.

Questions the method raises

What is generative art?+

Generative art is artwork produced through a system — algorithmic rules, machine models, or both — rather than drawn by hand. A human designs the process and its constraints; the system explores possibilities within them. Every image here comes out of a directed engine, not a blank prompt box.

How does Synthetic.pics create abstract art?+

The engine starts with a visual strategy, develops sixteen candidate art directions, and keeps the one that best balances coherence against novelty. A language model turns the winner into a concise image prompt, the image model renders it, and automated checks on prompt, title, description, and archive similarity run before publication.

How does the engine avoid repetitive images?+

Each candidate is compared against recent works across structural dimensions — strategy, form family, composition, surface, spatial relationships — with color weighted lower, so a new palette alone never counts as novelty. A text-embedding check catches reworded repeats, and removed works are remembered so the engine steers away from them. These checks reduce repetition; they work from direction descriptions, not pixels, so they cannot catch every visual repeat.

Is every image generated from a different seed unique?+

No. A seed makes a direction traceable and reproducible — the same seed under the same configuration and archive state gives the same direction — but different seeds can still converge on similar images, and different wordings can describe nearly identical compositions. Seed uniqueness is bookkeeping; the engine evaluates novelty structurally instead of trusting the seed.

What role does AI play in the process?+

Three roles. The art-direction logic — sampling, coherence scoring, novelty comparison — is deterministic software built by hand. A language model writes the image prompt from the winning direction, with a deterministic fallback. An image model renders the artwork. Above all three sits the human-defined system: strategies, vocabularies, compatibility rules, and validation gates.

Are all generated images published?+

No. Candidates that fail the coherence threshold are discarded before rendering; prompts that fail validation are rewritten or fall back to a deterministic generator; ungrounded titles or descriptions block publication. Survivors publish automatically. When a weak work slips through, we remove it — and the removal feeds back into the engine.

Practical questions about collecting, licensing, and the gallery are answered in the FAQ, including the rights associated with each work in licensing.

The archive

Explore the collection
GalleryCollectionsGenerative ArtAbstract ArtLicensingFAQ
  • compression
  • constellation-signs
  • spectral
  • charged
  • repeat-with-consequential-deviat
  • split-complementary
  • quiet