TheMindExpansionNetwork/mindbotz-style-sheets-v2

MINDBOTZ Style Sheets v2

20 professional character style/reference sheets, generated end-to-end by the Sonic Forage pipeline: an on-device prompt-expander LLM (Qwen3-VL inside Krea2) + Krea2 Turbo (int8) on a consumer RTX 4070, orchestrated remotely via SSH tunnel + ComfyUI API.

Every sheet follows a professional concept-art layout: character turnaround (front/side/¾), expression callout row, color palette swatch strip, props, clean background — each in a different art medium (oil painting, film noir poster, watercolor, pixel art, blueprint schematic, vaporwave airbrush, dieselpunk, risograph, art deco, sticker sheet, 70s van art, papercraft, stained glass, ukiyo-e woodblock, comic ink, chalk pastel, isometric game asset, kawaii anime, tarot card, vintage circus poster).

Files

Path	Contents
pairs/	Drag-and-drop ready: 20 × sheet_X.png + matching sheet_X.txt — drop the whole folder straight into Ostris AI-Toolkit / Kohya / ComfyUI trainers
images/	20 sheets, 1024×1024 PNG (sheet_<id>.png)
captions/	matching caption files (sheet_<id>.txt)
prompts/prompts.json	the 20 staged prompts (full reproducibility)
jobs.json	ComfyUI prompt_id ledger for every render
contact_sheet.png	5×4 preview grid of all 20 sheets
video/	64.8s slideshow with Ember soundtrack
scripts/	build/submit/harvest pipeline (the whole run is re-runnable)

Training use

The pairs/ folder is trainer-ready: image + same-named .txt caption side by side, standard LoRA convention. In AI-Toolkit's GUI, drag the folder onto the dataset picker and it ingests all 20 pairs. Caption style is single-sentence natural language (subject, medium, layout keywords, lowercase, comma-separated).

Prompt pipeline

1. Short human seed prompt (slug, medium, subject)
2. On-device LLM expands it with hard layout requirements: turnaround + expression row + palette strip
3. Krea2 Turbo, 8 steps, euler/simple, CFG 1, per-job fixed seed (90000 + 137·i)
4. Harvest + validation (size floor) + caption generation

Intended use

• Style-reference conditioning / image-prompting for character design
• LoRA training on "reference sheet layout" conventions
• Prompt-engineering study: how medium words steer layout consistency

Provenance

• Generator: Krea2 Turbo (int8 convrot) + krea2_style_reference LoRA, ComfyUI 0.34.1
• Hardware: NVIDIA RTX 4070 12GB (consumer, local)
• Date: 2026-08-27
• License: CC-BY-4.0 — the Sonic Forage studio empire strikes again

Known quirks (honesty section): view-to-view consistency of small accessories (hats) varies; expression rows lean subtle; garbled pseudo-text on labels is common in generated sheets.