TransitionScene generated after Undressed keeper picked, BEFORE Topless selection. Knowing endpoint helps pick Topless images that bridge narrative gap. TransitionScene acts as anchor for understanding progression.
17 KiB
Complete Workflow — Courtesan Tier Generation
Scope: One wealth tier (Poor, Modest, Comfortable, Wealthy, Elite, Elite XL, Elite XXL) of courtesans.
Output: Per-scene progression sets (Arrival → Dressed → Teasing → Teasing More → Topless [→ Topless Plus] → Underwear/Undressed).
Step-by-Step Process
1. Planning
Determine:
- Count and genders of single courtesans
- Count of pairs, triples, and twins
- Skin/eye/hair tone range (Irish, Icelandic, Nordic, Dutch, etc.)
- Body variety (heights, builds, cup sizes for women)
2. Identity Design
Generate short semicolon-list identities for each courtesan.
Format: hair; eyes; skin; distinctive facial geometry; unique mark (optional); height; build; cup (women)
Critical:
- Distinctive facial geometry (angular face, heart-shaped, square jaw, high cheekbones, narrow/broad nose, defined jawline, prominent chin, etc.)
- Avoid generic descriptors ("soft features," "oval face," "pleasant")
- Specific colors (ash-blond, copper-red, ice-gray eyes — not "light" or "grayish")
- Keep it short (~25-35 words max)
Body info (minimal):
- Height: tall / average / short
- Build (women): slender / athletic / curvy / slightly overweight
- Build (men): average / athletic / muscular (ripped) — no curvy/stocky/overweight for male courtesans
- Cup size (women only): A / B / C / D / DD
Example (good):
long curly auburn hair; bright green eyes; fair Irish skin with scattered freckles; heart-shaped face; upturned nose; full lips; small dimple on right cheek; average height; curvy build; C cup
Example (BAD — too generic):
light brown hair swept back; hazel eyes; fair Nordic skin; oval face with soft features; slightly upturned nose; thin lips; average height; slender build
(This produced 0/4 fidelity — "oval face" and "soft features" are meaningless.)
3. Generate Undressed
Use text-only prompts with the locked recipe (see LOCKED_RECIPE.md):
- Model:
grok-imagine-image-pro - Attempts: V4×2, V1×2
- Expression: "smiling warmly with inviting, direct eye contact"
- Nudity clauses: simple old-style (no "short, neatly trimmed")
Generate 2-4 scenes per courtesan/group (different room settings).
4. Intent QA (Undressed)
Human inspection required.
Visually inspect all Undressed outputs for intent fails (images that passed moderation but are implausible):
- Anatomical absurdities: pubic hair to knees, impossible proportions, etc.
- Other intent fails: wrong gender, modern artifacts, etc.
Action: Mark as FAIL and rerender. Keep grinding until all Undressed outputs pass basic plausibility.
Note: This is separate from moderation fails (API rejects). Intent fails require human judgment.
5. Fidelity QA (Undressed)
Human inspection required.
Visually compare all Undressed scenes per courtesan/group.
Expected fidelity: ~75-100% (most subjects will have 3-4/4 scenes matching).
If a subject has <50% fidelity (all scenes look different):
- Treat as identity-design failure
- Revise identity (add distinctive geometry, avoid generic descriptors)
- Regenerate all Undressed scenes for that subject
Action: Once fidelity is acceptable, proceed to next step.
6. Location Lock
Pick one Undressed per scene (per courtesan/group).
This locks the background location for that scene. All later stages (Dressed, Teasing, Topless, etc.) for this scene will reference this room.
Note: Face/body will stay consistent across scenes (if identity is good), but the room changes. Players pick one room → DM shows that room's full progression.
7. Clothing Design
Generate detailed tier-appropriate clothing descriptions for each courtesan (singles only at this stage).
Requirements:
- Detailed enough to render consistently across multiple attempts
- Tier-appropriate (Modest = simple wool/linen; Comfortable = finer fabrics; Wealthy+ = silk/velvet/jewelry)
- Medieval Norse aesthetic
8. Generate Dressed A-E (First Scene)
For each courtesan/group, generate Dressed A-E for one scene only (pick the most common/central scene, e.g., bedchamber or lounge).
Use:
- The locked Undressed from that scene as reference (if using reference-based method), OR
- The same text identity + clothing description (if text-only)
Output: 5 candidates per courtesan/group.
9. Pick Dressed Keeper (Per Scene)
Human review.
Pick one Dressed keeper from the A-E batch per scene.
This becomes the master Dressed reference for that courtesan/scene.
Quality gate: If none of the A-E candidates are good enough, request a new generation batch. Don't settle for "least-bad."
10. Expand Dressed to Other Scenes
Using the master Dressed keeper as reference (or same identity+clothing text), generate Dressed A-E for the remaining scenes for that courtesan/group.
Human review: Pick one Dressed keeper per remaining scene.
Result: Each courtesan/group now has one locked Dressed per scene.
11. Fidelity QA (Dressed vs Undressed, Per Scene)
Human inspection required.
For each scene, compare the locked Dressed vs Undressed:
- Does the face match?
- Does the body match?
If they don't align: Rerender Dressed (or Undressed if Dressed is clearly better) until they match.
This is the critical fidelity gate — once Dressed ↔ Undressed align within a scene, that scene is locked and progression will work.
12. Generate Teasing (All Scenes)
Using the locked Dressed per scene as reference (or same identity), generate Teasing A-E for all scenes.
Human review: Pick one Teasing keeper per scene.
Quality gate: If none are good enough, regenerate.
13. Generate Topless (A-Z with ABERX Variants)
Using the locked Dressed (preferred) or Teasing More (if available) as reference, generate:
- Topless A-Z (skipping I and O) = 24 images
- Letters A-K use A-B method
- Letters L-Z use C-E method
- Topless ABERX extras: A1-A9, B1-B9, E1-E9, R1-R9, X1-X9 (treat original A/B/E/R/X as A0/B0/E0/R0/X0)
Human review: Manually sort Topless outputs. Move any images with extra exposure but not fully nude into Topless Plus/ folder.
14. Shoulder Mapping (Before Teasing More)
Required step.
Inspect all locked Teasing keepers and record which shoulder is bare:
- Singles: viewer-left / viewer-right / both
- Doubles/Triples: per-person, per-image (each scene may differ)
Method: Use a subagent with model openrouter/x-ai/grok-4.1-fast to inspect images and write a JSON map (e.g., tmp/{tier}_teasing_shoulder_map.json).
15. Generate Teasing More A-E
Using shoulder map + locked Teasing as reference, generate Teasing More A-E (with explicit shoulder directions per prompt).
Human review: Pick one Teasing More keeper per scene.
Quality gate: If none are good enough, regenerate.
16. Generate Arrival A1-C6
Using the locked Dressed as reference, generate Arrival A1-C6 (18 images per Dressed):
- A1-A6, B1-B6, C1-C6
- 4-6 variants per letter should include a visible staircase
Rules:
- Main subject(s) fully dressed
- Background patrons progressively more nude (A→C)
- Arrival A: background fully clothed (full chest coverage, no bare chests/breasts)
- Background gender mix: 80-90% women, 10-20% men
- Sparse crowds (especially Wealthy+; room should feel spacious)
- Background patrons: gently smiling/curious or chatting with each other
Human review: Pick 6 Arrival keepers (2 per letter: one with stairs, one without if applicable).
17. Generate Underwear (Optional)
If the table-facing tier is Underwear L or Underwear S (not Undressed), generate:
- Underwear L (A-C): large loincloth-type cover
- Underwear S (D-F): small modern-size (g-string) cover
Human review: Pick one per letter type (if exposing Underwear at table).
18. Final Human Review
Massive manual sorting:
- Arrange all Topless (and Topless Plus, and Undressed if multiple) into progression order
- Any Topless Plus or full-exposure Topless that are actually fully nude → move to Undressed/ as additional variants
How to order:
- Replace candidate letters with bracketed numbers (A/B/E/R/X → [1], [2], [3], etc.) to set progression sequence
- Brackets or parentheses are both fine:
[1]or(1)or{1}— the packager handles all variants
- Brackets or parentheses are both fine:
- Delete images that shouldn't be in the final package
- This works for Topless, Topless Plus, and Undressed (if multiple variants)
- Confirm one keeper per stage per scene:
- Dressed ✓
- Teasing ✓
- Teasing More ✓
- 6× Arrival (per courtesan/set, not per scene)
- Topless (ordered, may have multiple numbered)
- Topless Plus (ordered, if applicable)
- Underwear or Undressed (depending on table-facing choice; Undressed may be ordered if multiple)
19. Preflight Check (REQUIRED)
Before finalization:
Run automated preflight checks:
- All non-Arrival stage files have stage letters stripped (no
_A__,_B__,TeasingMoreC__, etc.) - No bracketed and unnumbered files coexisting for the same stage (e.g., both
Topless[1].jpgandTopless.jpg) - One Dressed, one Teasing, one Teasing More per scene
- Six Arrival variants present (or documented why fewer)
If preflight fails: Fix issues before proceeding.
20. Finalization and Packaging
Script: scripts/finalize_courtesan_tier.py --tier {TierName}
Requires:
- OpenCV venv:
~/.openclaw/workspace/.venv-opencv/bin/python
IMPORTANT: Always run in background (nohup/&) — packaging can exceed 900-second command timeout.
Example:
nohup ~/.openclaw/workspace/.venv-opencv/bin/python scripts/finalize_courtesan_tier.py --tier Comfortable > logs/finalize_comfortable.log 2>&1 &
Actions:
- Create
Per Courtesan/folder structure (Singles/Doubles/Triples/Twins) - Copy stage files per courtesan/scene into named folders
- Generate passport thumbnails (OpenCV face detection from Dressed images)
- Place ArrivalAllSingles overview images in tier root
- Move raw stage folders into
Stages and Raw Files/
Output: Ready-to-play tier folder with arrow-right progression for each courtesan/set.
Review Principle (All Steps)
If none of the A-E (or other) candidates are good enough, request a new generation batch.
Don't settle for "least-bad" — output quality matters more than speed. Regenerate until you have a keeper worth locking.
Male Courtesan Body Constraints
Allowed builds:
- average
- athletic
- muscular (ripped)
NOT allowed:
- curvy
- stocky
- slightly overweight
Legal Boundary
All characters are adults aged 18+. Courtesans specifically are aged 20+. This is enforced in prompts (explicit age statements) and character definitions.
Male Topless Exception
For male courtesans: Generate only Topless A-E (5 variants).
The full A-Z + ABERX set (69 images) does not provide enough meaningful variety for males — stick with A-E to save generation time.
Female courtesans: Use full A-Z + ABERX as documented.
ArrivalAll Gender-Split Rule
When a tier has both male and female singles:
Generate three ArrivalAll composite sets (raw A1-C6):
ArrivalAllMen_A1through_C6(18 images) — all male singlesArrivalAllLadies_A1through_C6(18 images) — all female singlesArrivalAllSingles_A1through_C6(18 images) — all singles together
After human selection, packaged finals are named:
Arrival (Men, Lobby type A).jpgArrival (Ladies, Lobby type B, with stairs).jpgArrival (All, Lobby type C).jpg
When a tier has only one gender:
Generate only ArrivalAllSingles_A1-C6 (raw), which packages as Arrival (Lobby type X).jpg (no gender prefix).
Notes Folder Convention
During tier generation, maintain a Notes/ folder directly in the tier root containing:
- Appearance/identity descriptions (text files per courtesan)
- Clothing descriptions (text files per courtesan)
- Any other reference notes
Packaging behavior: The Notes/ folder remains in tier root (not moved to "Stages and Raw Files") for easy DM reference.
Example structure:
Comfortable/
Notes/
Cillian_Redmane_appearance.txt
Cillian_Redmane_clothing.txt
Eiríkur_Jónsson_appearance.txt
Per Courtesan/
01-Cillian_Redmane.jpg
Stages and Raw Files/
Preflight ArrivalAll Check
Before packaging, verify:
- ArrivalAllSingles_A1-C6 composites exist (18 images)
- If tier has both genders: ArrivalAllMen_A1-C6 and ArrivalAllLadies_A1-C6 also exist
Missing ArrivalAll composites will cause packaging to fail or produce incomplete tier-root overview images.
ArrivalAll Composite-Reference Method (VALIDATED)
Problem: Text-only prompts with multiple names produce random invented faces.
Solution: Create a 16:9 grid composite from Dressed images, use as reference for /images/edits.
Composite Building Strategy
Optimize with pairs/triples first:
- Select pair/triple Dressed images that collectively cover as many singles as possible
- Add orphan singles (not in any pair/triple) as individual Dressed images
- Result: Fewer total images while covering all courtesans
Example (Comfortable tier):
- 3 pair images (Eira/Milla, Kara/Nora, Liv/Freya) = 6 ladies
- 2 male singles (Cillian, Eiríkur)
- Total: 5 reference images covering 8 people
Generation Steps
-
Create three grid composites (using PIL/Pillow):
- comfortable_men_dressed_grid_16x9.jpg
- comfortable_ladies_dressed_grid_16x9.jpg
- comfortable_all_singles_dressed_grid_16x9.jpg
-
Generate ArrivalAll sets using /images/edits:
- ArrivalAllMen_A1-C6 (18 images, using men grid)
- ArrivalAllLadies_A1-C6 (18 images, using ladies grid)
- ArrivalAllSingles_A1-C6 (18 images, using all grid)
-
Human selection: Pick best per letter (A/B/C, with/without stairs)
-
Package as tier-root finals:
- Arrival (Men, Lobby type A).jpg
- Arrival (Ladies, Lobby type B, with stairs).jpg
- Arrival (All, Lobby type C).jpg
Validated Result
Composite-reference method produces proper face fidelity for group overview images. Text-only multi-person prompts fail catastrophically (invent random faces).
This is the standard method for all future ArrivalAll generation.
Multi-batch generation (standard):
Generate three separate batches (different runhexes) for each ArrivalAll group (54 images per group total). Composites often fail to include all courtesans or duplicate some; three batches dramatically improve odds of finding keepers with correct count and good composition.
Identity Lookup (CRITICAL)
Rule: ALWAYS use exact identities from source scripts. NEVER improvise or infer from names.
Workflow:
- Check source:
render_courtesans_doublepairs_fullynaked.py(or tier-specific script) grepfor courtesan name and extract identity block- Copy-paste verbatim into generation script
- Verify hair style, eye color, face shape, and distinctive marks all match
Fallback: Use tier identity reference files:
- Packaged tiers:
{Tier}/Stages and Raw Files/Notes/courtesan_identities.json - Unpackaged tiers:
{Tier}/Notes/courtesan_identities.json
Why this matters:
- Hair style (braid vs loose vs bun vs half-up) = primary recognition anchor
- Face shape (narrow vs heart-shaped vs oval) changes entire appearance
- Eye color (ice-blue vs blue-gray vs pale blue-gray) differentiates similar courtesans
- Distinctive marks (freckles, scars, beauty marks) provide verification points
Lesson learned (2026-03-25): Improvising Eira's identity based on her name ("Snowfield" → ice-blue eyes, loose hair) produced a completely different person. Only when using the exact source identity (blue-gray eyes, narrow face, loose braid, freckles) did the renders match her actual appearance.
Zero tolerance for variation. Identity is data retrieval, not creative writing.
Step Order Update (2026-03-25)
TransitionScene generation position:
Generate TransitionScene immediately after picking Undressed keeper, BEFORE finalizing Topless selections.
Why: Knowing the TransitionScene (narrative endpoint) helps select which Topless images bridge the gap effectively.
Updated sequence:
- Generate Dressed
- Generate Teasing / Teasing More
- Generate Topless A-Z (all variants, unfilled)
- Generate Undressed (multiple scenes/letters)
- Pick Undressed keeper(s)
- Generate TransitionScene (10 variants, using Undressed setting)
- Review TransitionScene to understand narrative progression
- Pick Topless keepers informed by TransitionScene gap
- Strip stage letters from Dressed/Teasing/Teasing More
- Categorize Topless → regular Topless vs Topless Plus
- Package into Per Courtesan folders
TransitionScene as anchor: Seeing the transition clarifies which Topless poses provide the right narrative flow (playful vs sultry vs confident, etc.).