Ekphrasis

ekphrasis

Classical ekphrasis is the vivid description of a visual work in words. That is exactly this: words engineered to produce an image, rather than words that describe a mood and hope.

Provenance

Distilled from the codeai-artist agent's own vault in Chatterbot 2000, written during a working conversation with the owner:

• AppData\Roaming\Chatterbot 2000\Vault\codeai-artist\ref-prompt-methodology.md
• AppData\Roaming\Chatterbot 2000\Vault\codeai-artist\ref-legibility-tell.md

⚠ This is a field method, not a paper. It was arrived at by generating images, looking at them, and naming what went wrong. Its claims about mechanism are stated below with their limits attached, because the method's own honesty is part of why it works.

The core stance

Slop is not cured by adding adjectives. Adding adjectives is how slop is made.

Slop comes from prompts that describe a mood instead of a subject. Given a mood, the model fills every unspecified gap with its most probable answer, and the sum of most-probable answers is the mean of everyone's averages. That mean is the glossy AI look.

The method is specificity plus constraint. It is the opposite of a style-word pile.

⛔ Delete on sight: 8k, hyperdetailed, masterpiece, award-winning, cinematic, stunning, breathtaking, ultra-realistic, trending on artstation. Every one of them pushes the sample toward the generic centre.

The six principles

1. Name a real subject with particularity. Not "a forest" but "a stand of birches in November, one leaning against its neighbour, bark peeling in narrow strips." Give the model something concrete to anchor to.

2. Commit to one light source with a physical direction. Slop lives in diffuse, directionless light. "Late northern light through a high window, long raking shadows" forces choices the model would otherwise average away.

3. Say what the image is NOT. Use the negative field: no lens flare, no chromatic aberration, no oversharpening, no soft-focus glow, no cinematic lighting, no 8k.

4. Restrict the medium explicitly. "Colour photograph" drifts to commercial gloss. "Shot on 35mm film, mild grain, halation" or "soft charcoal study, visible smudges and erased lines" fights the plastic sheen directly.

5. Give the composition a job. Ask for asymmetry, negative space, an off-centre subject, a quiet corner. Restraint reads as intention.

6. Know what it is FOR before writing. A cover, a mood board, a texture study and an editorial illustration each want a different prompt. Writing before knowing produces a prompt that serves none of them.

The legibility tell

The single most reliable giveaway, and most people cannot name it.

Image models are rewarded for text being readable. So they rotate documents, signs, books and labels flat-on to the viewer, at readable angles, in readable type, even when nobody in the scene would arrange them that way. The model is not reasoning about the scene's internal logic. It is arranging the world for the viewer's convenience.

Once you can see it you cannot unsee it. It is the same mechanism that makes AI hands hold objects at awkward camera-facing angles.

The four-move fix, strongest first

1. Take the text out of the image entirely. The strongest fix by a distance. "A neat stack of folded papers bound by a rubber band, lying flat." No text, no legibility pressure, no rotation problem. The meaning survives without visible words.
2. If text must be readable, make it partial and incidental. Never centred, never flat-on. "A worksheet lying askew at the table's edge, half under a saucer, the edge faded." Physics anchoring makes rotation harder.
3. Reduce the count. One legible element is a target. Three is a cluster the model arranges like a display. Cut to one, or swap readable things for unlabelled equivalents — a paper cup instead of a worksheet.
4. Add a recessional instruction. "The table is seen at a steep three-quarter angle receding into the scene; objects follow the table's perspective, none facing the viewer."

Inline negative that names the failure mode directly:

no readable documents facing the viewer, no text turned toward the camera, nothing arranged for the viewer's convenience — every object follows the scene's own perspective

⚠ Move 4 alone will not save you. Models obey it unreliably. Removal is the defence; the instruction is a bonus. When text is not in the frame at all, the tell has nothing to latch onto.

Animation: why sheets jitter, and the fix

The diagnosis: generators draw each frame autonomously with no shared coordinate system. Frame 3's hand does not belong to frame 2's clock. The jitter is the model having no consistent geometry across frames.

The fix: stop asking for frames. Ask for one numbered contact sheet with gutters. Inside a single image the model does hold geometry.

Lock everything static identical across frames. Change only the moving part, along a clean known sequence. End on a rest frame identical to frame 1 so it loops.

Phrases that do the actual engineering:

• "identical framing and identical camera angle"
• "the only thing that changes between frames is [the moving part]"
• "[static elements] remain absolutely identical in every frame"
• Move along a discrete known sequence (12, 3, 6, 9, 12), never "continuous motion"
• Give complex motion an arc: down, up, peak, fall, land, settle, rest
• For multi-limb motion, describe each frame as an explicit named pose, limb by limb

Assembly:

• Generate several times and pick the sheet whose static elements match best.
• Assemble at 6–8 fps, not 24. Stop-motion reads better slow, and slow hides inconsistency.
• To splice two sequences (walk into run): keep them as separate prompts sharing side view, character, medium and lighting. Both must begin and end on an identical standing pose. That pose is the seam. Hold it a beat at the join.

Honest limits

This is a strong tendency, not a law. A weak base model, anatomy collapse, or the platform's own post-processing can reintroduce slop no matter how good the prompt is.

The real mechanism is distribution averaging: vague prompts push samples toward high-probability regions of the training distribution. Models synthesise from a learned distribution. ⛔ They do not retrieve stored images, and any explanation that assumes they do will mislead you about what to change.

Expect to steer conversationally after the first render — "turn the stack so I see only its edge", "push the jar further into shadow", "make the table slope away more steeply". A first generic result is normal, not failure.

Working method

1. State what it is FOR in one sentence.        → principle 6
2. Name the subject with particularity.         → principle 1
3. Fix the light: one source, one direction.    → principle 2
4. Fix the medium.                              → principle 4
5. Give the composition a job.                  → principle 5
6. Strip text from the frame.                   → legibility tell, move 1
7. Write the negative line.                     → principle 3 + the tell
8. Generate several. Pick. Then steer in words.

Composes with

• image — this writes the prompt; that generates, edits and optimises the resulting file. Use both.
• hyperframes / music-to-video — assemble contact sheets into motion.
• peitho — the prose counterpart. Same shape of problem: both are about refusing the most probable phrasing.
• ad-creative — platform specs and sizes once the art exists.

Deliberon's codeai_artist specialist (specialist_send) can author prompts with this method, and specialist_gfx_generate renders them. ⚠ Both need Deliberon running.