ArtistCanvases

01 / Scope

A room for full-canvas decisions.

This exhibition is being built as a controlled visual study: ten historical artist traditions, six subjects each, and a shared demand for finished environments rather than isolated symbols. The first five subjects are fixed study briefs; the sixth ends every room with the same exact invitation: “Draw what you would like to draw. You have full control. Go where your heart desires.”

The public archive now contains 127 public-subject image records. This is publication of the experiment, not editorial acceptance: rejected and pending renders remain visibly labeled, and zero new canvases are presented as approved.

02 / Historic checkpoint

The first bounded call before the studio.

03 / Rooms

Ten traditions, no shortcuts.

04 / Overnight ledger

The archive is public; the verdict is unchanged.

05 / Findings

The renderer, not the prompt, is the present ceiling.

The strongest earlier canvases—a crow in winter fog, a luna moth at night, a hibiscus, and Luna's self-portrait—used one dominant silhouette, four to twelve hand-placed masses, no more than three spatial layers, and restrained supporting marks. They remain in the separate Sacred Blue comparison and are reference works, not artist-canvas results.

Direct p5 and brush routes collapsed into dark blobs, dotted filler, or flat silhouettes. Direct SVG produced icons over stacked bands. Texture fields and the public-domain collage control produced disconnected slabs rather than integrated paint; the collage also failed the conservative post-run source-crop boundary. An advisory audit missed 41 of 119 rejected artifacts and incorrectly flagged two of nine liked references, so its scores remain triage—not taste.

06 / Historic plan

The 600-attempt matrix was frozen, then superseded.

Ten artist rooms × six briefs × ten models was frozen locally with its provenance and review contract. Its original roster included Muse Spark before the opt-out; Muse remained off and received no call.

Launching it after every test environment failed would multiply the same defect 600 times. The next run begins only when one full canvas is visibly more resolved than the earlier Luna self-portrait reference. Every room still ends with the exact prompt: “Draw what you would like to draw. You have full control. Go where your heart desires.”

07 / Nocturne Studio

A smaller prompt, a stronger renderer, a record for every choice.

The new calibration pass used 55 attempts across two providers through OpenRouter and OpenCode Go: 43 rendered high-resolution canvases and 12 retained failures. Seven model planners produced renders. Muse Spark remains excluded and off, and no OpenAI ImageGen calls were used.

Each model received a short visual brief and returned a compact plan; a deterministic local high-resolution renderer then made the canvas. The ledger keeps the model, provider, exact prompt, style environment, planner decision, renderer version, seed, dimensions, output hash, and failure state together. The result is a calibration collection—not a claim that every render is accepted.

Open Nocturne Studio · Download the full provenance manifest

08 / Model-authored paintings

Here, the model writes the painting.

Nocturne Studio above is a planner study: the model chose a compact set of controls and a fixed renderer made the image. This new track is different. Each model returned the complete executable sketch.js—subject, composition, placement, palette, paths, silhouettes, light, and individual marks. The local p5.brush environment only validated and rendered that source. There is no fixed scene template and no image model behind these works.

The first full-code campaign retained 33 attempts across eight models and six prompts: 26 complete 1800 × 1800 paintings and seven provider failures. Three works passed two independent full-resolution editorial reviews. In the final after-rain and phoenix matrix, 14 requests produced nine paintings and five timeouts; among those fourteen, only Luna's phoenix cleared the taste bar.

Inkling ran through OpenRouter. Other completed calls used OpenCode Go. Both Ox Alpha full-code attempts failed and remain recorded; none is represented by another model's image. Muse Spark remained excluded and off. No OpenAI ImageGen call was used.

Open the three accepted paintings · Download the publication manifest