Mixmatter

AI art direction and source-aware visual reconstruction for existing imagery.

Mixmatter treats photographs as visual matter: something to read, protect, cut, flatten, layer, and rebuild with intent.

Mixmatter is an open-source Agent Skill / SKILL.md that reads a source image, decides what must survive, forms a source-specific art-direction thesis, and reconstructs the image into a contemporary print-driven composition.

Photography is source material, not sacred material.

Reconstruct, do not decorate.

Preserve semantic identity, not visual completeness.

Mixmatter 2.1 keeps the visual DNA established in v1 while adding stronger source understanding, preservation logic, art-direction judgment, critique, and revision continuity. It remains a Skills-only system. There is no required MCP server or custom ChatGPT host UI.

Install

npx skills add Fanjiale-CN/press-print --list
npx skills add Fanjiale-CN/press-print --skill mixmatter

Repository: Fanjiale-CN/press-print
Skill: mixmatter
Entry point: SKILL.md

Identity: the OpenAI package and Skill slug are mixmatter. The current GitHub repository URL remains Fanjiale-CN/press-print until the repository slug itself is renamed.

How Mixmatter 2.1 works

The internal reasoning sequence is:

READ
→ UNDERSTAND
→ PROTECT
→ DIRECT
→ RECONSTRUCT
→ MATERIALIZE
→ CRITIQUE
→ REVISE

The user does not need to operate those stages manually. A clear request executes directly. A vague request is inspected and art-directed autonomously unless genuine ambiguity would materially change the result.

READ / UNDERSTAND

Identify semantic anchors, visual anchors, structural relations, identity invariants, clutter, low-information fields, directional forces, and transformation opportunities.

PROTECT

Classify source information into:

must preserve
should preserve
may transform
may remove

Explicit user locks always outrank defaults.

DIRECT

Form one dominant source-specific thesis about what to amplify, suppress, reframe, flatten, isolate, or fragment.

Direction is structural judgment, not a generic style preset.

RECONSTRUCT

Prefer structural operations before effects:

crop / reframe
isolate
suppress
scale contrast
planar compression
controlled fragmentation, overlap, or repetition when justified

A successful default result should be visibly reconstructed, not merely photo-styled.

MATERIALIZE

Use halftone, duotone, torn/cut edges, paper layering, registration error, photocopy behavior, and related print materiality only when they reinforce the composition.

CRITIQUE / REVISE

Preserve successful decisions and fix the cause of a failure rather than re-randomizing the whole image.

Natural-language controls

Mixmatter retains three useful semantic controls without requiring a custom UI:

Direction: the dominant reconstruction thesis
Structure: how much original compositional continuity may be spent
Intensity: how strongly the chosen treatment appears

Users may state these explicitly or simply describe what they want in normal language.

Core visual identity

Mixmatter generally prefers:

flat, surface-first composition
compressed or interrupted depth
strong editorial hierarchy
source-derived crop and scale shifts
interlocking graphic planes
selective photographic retention
visible but selective halftone / duotone
bold silhouette and shape logic
controlled tactile collage
active low-information fields
contemporary print energy rather than generic retro styling

It avoids:

intact photo + cosmetic texture
generic premium-ad polish
cinematic realism drift
blanket halftone
uniform torn-paper treatment
arbitrary fragmentation
generic vector tracing
culture-as-costume styling
template sameness across unrelated sources

Source text policy

Mixmatter adds zero new text by default.

If source text already exists, it may be preserved, cropped, obscured, fragmented, reduced, or treated as source imagery. Identity-critical wording should be preserved faithfully when feasible.

Mixmatter must not:

invent approximate replacement wording
translate source text by default
create bilingual duplicates by default
turn source signage into a newly typeset headline without instruction
hallucinate pseudo-text when exact text cannot be reproduced

If the user supplies exact wording, use only that wording unless they explicitly authorize generated copy.

Research-derived system

Mixmatter 2.1 is grounded in seven canonical system documents:

These preserve the research layer independently of any particular interface implementation.

Visual authority

The established v1 rendering language remains available at:

The 2.0 reasoning layer extends that visual language rather than replacing it.

Showcase

The repository retains the ten canonical high-resolution v1 source/result showcase plates in examples/showcase/. See examples/README.md for notes and evaluation guidance.

Evaluation

Use:

The 2.0 benchmark evaluates semantic preservation, Mixmatter identity, reconstruction strength, editorial hierarchy, planar coherence, material coherence, source specificity, and common drift penalties.

Repository structure

mixmatter/
├── SKILL.md
├── skills/mixmatter/
│   ├── SKILL.md
│   └── references/
│       ├── mixmatter-v1.md
│       ├── mixmatter-v2-runtime.md
│       └── quality-rubric.md
├── docs/
│   ├── MIXMATTER_2_RUNTIME.md
│   └── system/
│       ├── MIXMATTER_ART_DIRECTION_POLICY.md
│       ├── MIXMATTER_CORE_CONSTITUTION.md
│       ├── MIXMATTER_REGRESSION_BENCHMARK.md
│       ├── MIXMATTER_SYSTEM_SCHEMA.yaml
│       ├── MIXMATTER_TOOL_SPEC.md
│       └── MIXMATTER_VISUAL_GRAMMAR.md
├── prompt/mixmatter-v1.md
├── eval/
└── examples/

Current status

Version: 2.1.0
Architecture: Skills-only, UI-independent
Status: Research-integrated release candidate

Author

Created by Fan Jiale / Galok.

License

MIT for the Mixmatter skill text, prompt system, documentation, and related project materials. See LICENSE.

Showcase image rights may depend on their original provenance. See examples/README.md.