Self-Portrait
Modes

The record

The first broad campaign gave each model route both instructions. Each author chose its own concept, words, palette, canvas, style, tools, and executable painting source. A blind visual critic reviewed the rendered canvas; the author then saw the image, revised it, and faced a final jury.

The second edition adds sharply different premises—split listening, weather, aperture, fracture, handwritten atmosphere, and typographic counterpoint—without sharing a composition template. The tools now support bounded pigment, dry-brush, and cut-paper lettering so language can enter the painted surface instead of becoming a clean title pasted above it.

A later unfiltered-transmission suite supplies no sentence, topic, slogan, or emotional conclusion. Each model chooses what it wants to put before a human, then paints one to four sentences as interrupted memory, crossed inner voices, erasure, contradiction, repetition, or an impossible question. The exact authored language is preserved with every work.

The newest word-built-face suite is stricter: white ground, small black single words, and no brush or geometric mark at all. Each model authors its own vocabulary and placement system, but every accepted source must produce at least 600 audited word calls at runtime. The strongest canvases use thousands. Their language forms one uncanny private transmission—watching, names, doors, permission, errors, memory, repetition, and an echo answering itself—while density and absence assemble a cropped face.

Two intensified fractured-face waves then commissioned sixteen independent canvases across eleven routes. Eight produced valid word-only renders and five were selected. The new direction permits doubled eyes, displaced mouths, split profiles, interrupted contours, and a second face leaking through the first, but still supplies no coordinates or common geometry. The authors chose their own transmissions: a name corrected during sleep, a witness rewriting the speaker, a receiver keeping an earlier voice, a voice behind a locked room, and a name that refuses to stay.

The two works that established this direction are preserved exactly as selected. New variants were generated independently rather than derived from a shared facial template; a content-aware crop removes accidental dead margin while retaining a narrow white border.

The selected exhibition is a supervisor-curated view of the strongest completed canvases. Automatic jury decisions remain visible as provenance, not as a substitute for visual judgment.

The frontier rooms ask the model to author a complete document: HTML and CSS compositions, inline SVG printmaking, or a raster canvas with its own material grammar. There is no supplied image template. Sources are rendered locally with network access blocked, then inspected at full resolution by a supervisor; a valid render is not automatically a good one.

Frontier works without automated critic scores say so explicitly in the lightbox. Their record still preserves the exact prompt, renderer, generation identifier when available, observed cost, hashes, and provenance link, while the archive keeps every completed image—including beautiful experiments that did not make the selected wall.

Muse was pending opt-in during this run and Ox Alpha was absent from the live catalog, so neither is represented as a completed route.

Every completed work carries its model, provider, mode, exact prompt, words, style, tools, canvas, attempt ID, source and image hashes, selection route, and curation rationale when selected. Jury decisions and scores remain attached where that evaluation stage ran; newer supervisor-only rooms are labeled honestly instead. The separate archive shows every image-bearing completion, including works not selected; true no-image failures remain excluded.