This LoRA was trained with a modular, atomic tagging structure designed for flexible prompt-building rather than fixed per-image descriptions.

Trigger word: horror_dreamcore

Shared vocabulary reference (identical wording wherever a feature repeats):

• Helmets: full-face helmet, bubble helmet, dome helmet, chrome visor helmet, tinted visor, goggles

• Materials: wet-look finish, chrome finish, latex texture, translucent panel, worn texture, weathered texture, splatter texture, mesh texture, iridescent finish

• Structure: bodysuit, armor plating, robe, trench coat, cargo pants, hooded robe

• Features: exposed midriff, exposed organs, exposed muscle, exposed ribs, tentacle growths, mushroom growths, crystal growths, foliage embedded, vine circuitry, glowing veins, glowing orbs, glowing accent dots, tube wiring, mechanical joints

• Armor parts: shoulder pauldron, chest plate, chest window panel, hip armor, knee armor, ankle armor, gauntlets, boots, sandals, backpack device, satchel pouch, utility belt, strap harness

Tagging approach:

• Every caption uses short, isolated natural-language phrases (materials, colors, structure, and features each get their own tag rather than being fused together).

• Repeated concepts use identical wording across the entire dataset (e.g. "dome helmet" always means the same silhouette, never swapped with "bubble helmet" or "glass dome" for the same feature) so the model reinforces one strong concept instead of several weak ones.

• Constant elements shared across the whole dataset (multi-view sheets, studio backdrops, standing poses) were deliberately pruned from every caption. These traits are absorbed directly into the trigger word, so horror_dreamcorealone reproduces that consistent presentation style without needing to spell it out in your prompt.

What this means for prompting:

• Call horror_dreamcore for the baseline dreamcore-horror armor/character aesthetic.

• Add modular descriptor tags (helmet type, material, color, mutation/growth feature, armor part) to pull specific traits from different trained designs and recombine them into new characters.

• Keep tag stacks short (4–6 modular tags after the trigger) for the cleanest results — over-stacking conflicting structure tags (e.g. two competing garment types) will fight each other in generation.