Keeping an AI-generated slot set consistent

Updated 2026-08-16

A folder of individually good images is not a set

Ask an image model for ten objects in the same style and it may give you ten plausible interpretations of the phrase “same style.” One has soft backlight, another a hard upper-right key. One uses painted stone, another glossy plastic. The hero fills 90% of its square while a rank occupies half. Worse, the signature character can leak into every prop because the full theme was repeated in every prompt.

Consistency is therefore built as a sequence. Each stage controls a different source of drift: visual reference, textual contract, subject isolation, measured framing and final colour correction. Remove one and another mechanism may hide the failure in a static contact sheet, but the moving reel exposes it.

1. Begin with a subject-free style board

The first generated image is a 3×3 chart of colour swatches, lit spheres and material chips on a neutral field. It contains no character, emblem, location or story object. That absence is the important design choice.

A reference image transfers more than intended. When an early pipeline used the Wild as the only style reference, a fast image model propagated the Wild's silhouette across other symbols. The renderer had interpreted “match this” correctly but had no reliable boundary between rendering and subject. A chart provides material, colour, light, value and edge evidence without supplying a hero shape to copy.

When the palette field is blank, the board alone receives the world text and chooses three colours suited to it. Every later reference-capable call sees the resulting chart, so the colour decision is made once. The code does not invent a hue from a theme dictionary. When a route cannot accept reference images, it runs on the text lock and is expected to drift more.

2. Compile one numeric style contract

“Polished premium 3D” is not a sufficient constraint. The pipeline compiles the selected style and palette into the same byte-identical block for every object asset:

• Dead-on orthographic camera, optical centre, no tilt, foreshortening or rotation.
• One key light from the upper left at 35 degrees, with 15% ambient fill from the lower right.
• Exactly three value bands per material plus one specular highlight cluster.
• A named material vocabulary and edge treatment for the chosen style.
• An approximate 60/30/10 dominant, secondary and accent colour split.
• No bloom, vignette, chromatic aberration, grain, lens flare, motion blur or depth-of-field blur.

The contract stays unchanged even where an asset needs a camera exception. The mascot prompt states its three-quarter body turn separately instead of silently altering the shared block. That makes accidental prompt drift visible and preserves the common renderer, light and material rules.

3. Make one symbol the approved master

The Wild is generated first and defines the set's real art baseline. It is measured against structural checks but not against the style-board average, because a chart of swatches on grey is not expected to share the value distribution of finished art.

Compatible models can receive both the style board and the approved master. The board says how the set is rendered without a subject. The master shows what the rendering looks like when applied to a production symbol: real contrast, edge density, material depth and reel-scale focus. The model must be able to follow the instruction to copy rendering rather than content; routes that cannot reliably separate those roles receive only the board.

The mascot has one deliberate content reference. Its finished full-height art is cropped to a head-and-shoulders sheet and sent to the mascot's reel bust. There, copying the same face is the goal. The distinction between a style reference and an identity reference is what prevents a reference system from becoming a motif-duplication system. The mascot design guide explains that exception.

4. Permit one subject and exclude the rest

The theme sentence does not appear in ordinary high- or low-tier prompts. Only the Wild and Scatter receive the world description because they are thematic emblems. Every other prompt names its one permitted subject and explicitly excludes motifs owned by other cast members.

This solves two opposite problems. The symbols retain the same rendering, but they do not become the same object. Loose atmosphere such as weather, particles, scenery and story is banned as a category and assigned to the background. It is not repeated by name under “do not,” because mentioning a noun can raise the chance that the model draws it.

Anatomy follows the same rule. Props receive exclusions for faces, eyes, hands, wings and figures. Only subjects the user marked as characters are permitted anatomy. A consistent set is not one where every kettle, key and rank has the mascot's eyes.

5. Keep the tier budgets ordered

Rendering consistency must leave room for hierarchy. The Wild and Scatter use 10–13 shape masses, the high tier 8–11 and the low tier 3–5. Premium richness comes from material layering and chamfer depth, while the low tier keeps one material and no ornament. Separate plaques appear only under special and high symbols.

Those controlled differences are part of consistency, not exceptions to it. A set can share light and palette while varying visual weight. The full method is in slot symbol tiers and paytable art.

6. Normalize what prompts cannot guarantee

A prompt can ask for a 12% margin and still receive a subject that fills 98% of the canvas. The pipeline therefore removes the backdrop, measures the alpha bounds, and reframes every object onto a fixed output canvas. Special, high and low symbols are normalized to 80%, 76% and 70% fill on 512×512 files.

Generated plaques are keyed and framed onto that same canvas. Glow, shadow, outline and rim are derived from the final symbol alpha, so every layer shares an origin without relying on the model to repeat a silhouette. The cabinet is uniformly fitted to a published rectangle, then its real dark panel is measured so the 5:3 reel opening follows the art.

7. Harmonize from the finished set

After framing, the pipeline builds merged palette anchors from the set's own finished art. Each symbol pixel is converted into LAB colour, matched to the nearest anchor, and pulled toward it in chroma while lightness is preserved. The default harmonization strength is 0.5; zero disables it, while a stronger pull such as 0.8 can flatten the identity of an individual object.

The anchors do not come from the style board. The board's neutral field would dominate its measured palette and drain colour from the symbols. Harmonization must follow the actual art it is correcting. It also does not repaint shadows or rebuild shapes: preserving L* keeps the form and value structure intact.

8. Measure once, never regenerate silently

Post-processed assets are checked for palette drift, mean lightness, excess detail, duplicate silhouettes, fill, edge contact and backdrop coverage. Those measurements are diagnostic. They do not trigger another paid call and they are not shown to the customer as a grade. A valid image is kept until the user chooses a redraw.

A redraw replaces only the chosen generated source and then replays harmonization, registered layers, manifest and previews. Its per-asset note sits beneath the original constraints instead of opening a second prompt path. For good input to that system, see how to write a slot art brief.

Review consistency in the right order

1. Check that every subject is different before judging shared style.
2. Compare upper-left light, material vocabulary, edges and three-band value structure.
3. Reduce symbols to 96 pixels and verify that tier differences survive.
4. Overlay registered layers and confirm no scaling or origin shift.
5. Spin the grid; duplicated silhouettes and noisy edges are easier to see in motion.
6. Compare desktop and portrait as the same location under different compositions.

This process is where a pipeline can outperform an unstructured chat session: it remembers which decisions belong to the whole set and which belong to one asset. It does not remove the need for taste or revision. The honest tradeoffs against a commissioned team are covered in AI slot art versus an outsourced studio.