Keeping an AI-generated slot set consistent

Updated 2026-08-24

Individually strong images do not automatically form a set

An image model can interpret "the same style" differently for each object: soft backlight on one, a hard upper-right key on another, or inconsistent materials and scale. Repeating the full theme can also spread one character or emblem across unrelated props. A consistent set therefore needs a visual reference, fixed text contract, subject isolation, measured framing and final colour correction.

1. Begin with a subject-free style board

The first 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. This preserves material, colour, light, value and edge evidence without giving later calls a hero shape to copy. An earlier Wild-only reference caused the Wild silhouette to spread through the set.

When the palette is blank, only the board receives the world text and chooses three suitable colours. Every reference-capable call then receives the same chart. The code does not select hues from a theme dictionary. Routes without image-reference support use the text lock and may drift more.

2. Compile one numeric style contract

The selected style and palette compile into one byte-identical block used by every object prompt:

• 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.

Asset exceptions do not alter the contract. The mascot adds its three-quarter body turn separately, preserving the common renderer, light and material rules.

3. Make one symbol the approved master

The Wild is generated first and sets the production baseline. Structural checks apply, but its values are not compared with the grey style-board average. Compatible models receive both the board and approved master: the board defines rendering, while the master shows production contrast, edge density, material depth and reel-scale focus. Routes that cannot separate style from content receive only the board.

The mascot is the identity-reference exception. Its full-height art is cropped to a head-and-shoulders sheet for the reel bust, where copying the same face is required. See the mascot design guide.

4. Permit one subject and exclude the rest

Ordinary high- and low-tier prompts omit the theme sentence. Only Wild and Scatter receive the world description. Every other prompt allows one subject and excludes motifs assigned to other symbols. Weather, particles, scenery and story remain background categories rather than named negative-prompt nouns, because mentioning a noun can encourage the model to draw it.

Prop prompts also exclude faces, eyes, hands, wings and figures. Anatomy is allowed only for user-marked characters.

5. Keep the tier budgets ordered

Wild and Scatter use 10–13 shape masses, high tier 8–11 and low tier 2–4 with no small detail clusters. Premium tiers use material layering and chamfer depth; low tier uses one material and no ornament. Plaques appear only under special and high symbols. These controlled differences preserve hierarchy within shared light and palette. See slot symbol tiers and paytable art.

6. Normalize what prompts cannot guarantee

Prompts cannot guarantee margins, so the pipeline removes the backdrop, measures alpha and reframes each 512×512 symbol. Special, high and low tiers are normalized to 80%, 76% and 70% fill. Generated plaques are keyed onto the same canvas; glow, shadow, outline and rim derive from final alpha and share its origin. The cabinet is uniformly fitted, then its dark panel is measured for the 5:3 reel opening.

7. Harmonize from the finished set

After framing, merged palette anchors come from the finished set. Each symbol pixel is converted to LAB, matched to the nearest anchor and pulled toward it in chroma while preserving lightness. Strength 0 disables harmonization, 0.5 is the default, and 0.8 can flatten an object's identity.

Anchors do not come from the style board because its neutral field would dominate the palette. Preserving L* retains form, value structure and shadows.

8. Measure once, never regenerate silently

Post-processed assets are measured for palette drift, mean lightness, excess detail, duplicate silhouettes, fill, edge contact and backdrop coverage. These diagnostics neither trigger paid regeneration nor appear as customer grades. Valid art remains until the user requests a redraw.

A redraw replaces only the chosen generated source, then replays harmonization, registered layers, manifest and previews. Its asset note follows the original constraints rather than creating another prompt path. See how to write a slot art prompt.

Review consistency in the right order

1. Confirm that every subject is distinct before judging shared style.
2. Compare upper-left light, materials, edges and three-band values.
3. Reduce symbols to 96 pixels and verify tier separation.
4. Overlay registered layers and check scale and origin.
5. Spin the grid to expose duplicate silhouettes and noisy edges.
6. Compare desktop and portrait as compositions of one location.

This process separates set-wide controls from asset-specific decisions. It supports consistent AI slot art but does not replace review or revision.