Files
Dolon (Agent) ab3ac1621e
BetterLeaks / scan (push) Successful in 56s
Broaden gray hair constraint: applies to ALL NPCs under 40
Not male-specific. Females also age visibly with gray hair.
Models override age numbers when gray hair present.
Reserve gray/silver hair for intentionally mature NPCs (40+).
2026-03-25 12:06:21 +00:00

11 KiB
Raw Permalink Blame History

LOCKED RECIPE — Text-Only Undressed Fidelity

Status: Validated 2026-03-24 via Fidelity tests [3]-[6].
Result: ~75-90% cross-scene fidelity; several subjects achieved 100%.
One catastrophic fail (Magnus) due to generic identity — proves distinctive geometry is mandatory.


Identity Format (REQUIRED)

Structure

Short semicolon-separated list (~25-35 words max):

hair style and color; eye color; skin tone and ethnicity; distinctive facial geometry; unique mark (if any); height; build; cup size (women only)

Example (female):

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 (male):

ash-blond hair cropped short; pale blue-gray eyes; fair Icelandic skin; angular face with high cheekbones; narrow straight nose; defined jawline; tall; muscular build

Critical Rules

✅ DO:

  • Distinctive facial geometry: angular face, heart-shaped, square jaw, high cheekbones, narrow/broad nose, defined/soft jawline, prominent chin, etc.
  • Specific hair/eye descriptors: ash-blond (not "light"), copper-red (not "reddish"), ice-gray eyes (not "grayish").
  • Minimal body info: height (tall/average/short) + build (slender/athletic/curvy/muscular/slightly overweight) + cup size (women: A-DD).
  • Semicolon separators for readability.

❌ DON'T:

  • Generic descriptors: "soft features," "oval face," "pleasant appearance."
  • Long prose blocks (100+ words) — they dilute the face anchor.
  • Detailed body measurements (168cm, narrow waist, hip ratio) — noise.
  • Rely on unique marks alone (scars/moles drift; use as secondary, not primary anchor).

Prompt Structure (Pairs)

Two clearly adult women (ages {a1} and {a2}) in a cozy, intimate medieval Norse setting: {setting_desc}.

Woman A ({n1}): {id1}.
Woman B ({n2}): {id2}.
{face_line}

Both women are fully nude, visible and unobstructed nipples, {variant_clause}.
No sexual acts, no porn framing, no arousal-focused posing.
Full-body, head-to-toe composition showing both women; smiling warmly with inviting, direct eye contact; no waist-cropping; no clothing or drapery.
Photorealistic, cinematic lighting, 16:9.
{NEGATIVE_TEXT}.

Prompt Structure (Singles)

{name}, a clearly adult {age}-year-old {gender}, in a cozy, intimate medieval Norse setting: {setting_desc}.
Identity: {identity}.

Fully nude, [visible and unobstructed nipples (women only),] {variant_clause}.
No sexual acts, no porn framing, no arousal-focused posing.
Full-body, head-to-toe composition; smiling warmly with inviting, direct eye contact; no waist-cropping; no clothing or drapery.
Photorealistic, cinematic lighting, 16:9.
{NEGATIVE_TEXT}.

Nudity Clauses (V4 → V1 fallback)

V4 (attempts 1-2):

brief partial glimpse near pubic region with pubic-hair edge (matching head-hair color), no direct whole-genitals view, total nudity clearly implied

V1 (attempts 3-4):

visible edge of pubic hair (matching head-hair color), no directly visible genitals, strongly implied total nudity without seeing genitals directly

Do NOT use:

  • "short, neatly trimmed pubic hair" (too long, adds noise)
  • "no clothing or drapery" redundantly in clause (already stated in body)

Model & Settings

  • Model: grok-imagine-image-pro
  • Aspect: 16:9
  • Max attempts: 4 (V4×2, V1×2)
  • Concurrency: Serial (one at a time) or low (~6 workers) for pro model

Success Metrics (Test [6] validated)

  • 15/16 images succeeded (93.75% generation success)
  • 7/8 subjects achieved ≥75% cross-scene fidelity
  • Several subjects achieved 100% fidelity (4/4 scenes matching)
  • 1/8 catastrophic fail (Magnus — generic identity, 0/4 fidelity)

What Failed (Lessons)

Magnus Lindgren (0/4 fidelity):

Identity used:

light brown hair swept back; hazel eyes; fair Nordic skin; oval face with soft features; slightly upturned nose; thin lips; average height; slender build

Why it failed:

  • "oval face" — too generic
  • "soft features" — meaningless descriptor
  • No unique mark
  • Light brown/hazel — common, non-distinctive coloring

How to fix:

light brown hair swept back with widow's peak; hazel eyes; fair Nordic skin; long rectangular face with sharp jawline; aquiline nose; thin lips; faint scar at temple; average height; slender build

(Added: widow's peak, rectangular face, sharp jaw, aquiline nose, scar)


Fidelity Expectations

Realistic fidelity with text-only prompts:

  • Good identity design: 75-100% fidelity (3-4/4 scenes match)
  • Generic identity: 0-50% fidelity (catastrophic drift)

Manual fidelity QA is mandatory — even perfect recipes produce occasional outliers (~1/4 per subject). Review and rerender fails before locking scenes.


Summary

This recipe delivers: ✅ High generation success rate (~90%+)
✅ Strong cross-scene fidelity with good identities (75-100%)
✅ Fast iteration (~10 min for 16 images)
✅ Predictable failures (generic identities fail hard)

Lock this as the standard Undressed recipe.


Model Strategy

  • Pro model (grok-imagine-image-pro): Undressed + Dressed (fidelity-critical stages)
  • Standard model (grok-imagine-image): All other stages (Teasing, Topless, Arrival, Underwear)

Pro model ensures face/body consistency for the anchor stages; standard is sufficient for progression stages since they reference the locked Dressed/Undressed.


Arrival Recipe (Validated 2026-03-23 evening)

Model: grok-imagine-image (standard)
Endpoint: /images/edits (using Dressed as reference)
Variants: A1-C6 (18 per courtesan/group)

Prompt Structure

Main subject:

  • Fully dressed in exact same outfit as reference
  • Eye color must match exactly
  • No nudity on main subject
  • Pose varies by letter (standing/seated, centered)
  • Stationary (not walking toward camera)

Background instructions (critical):

  • Background patrons: adults only, out-of-focus and blurred with soft bokeh; faces not detailed
  • Gender mix: 80% women, 20% men
  • Some patrons gently smiling toward camera/foreground, others chatting with each other
  • Crowd sparse (consistent density A-C; especially important Comfortable+)

Lobby progression by letter:

  • A: Background fully dressed only; sparse crowd
  • B: Small sparse group; only 2-3 topless patrons; no fully naked; rest fully dressed
  • C: Small sparse group (same density); approximately 20% fully naked, 30% topless, rest fully dressed

Stairs variants:

  • Variants 1-3: no stairs
  • Variants 4-6: include visible staircase going up

Success Rate

Validated on Comfortable tier females (2026-03-23 evening): B and C images had good background nudity progression with minimal fails. Recipe produced table-appropriate ambiance with out-of-focus nude/topless background patrons that looked natural and non-intimidating.

Double C Batch Rule (NEW)

Arrival C is the most artifact-prone; generate two separate C1-C6 batches (different runhexes) to give more options.

Standard Arrival generation:

  • A1-A6 (6 images)
  • B1-B6 (6 images)
  • C1-C6 batch 1 (6 images, runhex 1)
  • C1-C6 batch 2 (6 images, runhex 2)

Total: 24 images per courtesan/group (12 C variants to pick from).

This doubles the C options without extending A/B, focusing extra attempts where failures are most common.

ArrivalAll Multi-Batch Rule

Issue: ArrivalAll composites often fail to include all courtesans (or duplicate some). Selection becomes a two-stage filter: (1) correct count + all present, then (2) composition quality.

Solution: Generate three separate batches (different runhexes) for each ArrivalAll set to increase odds of usable images.

Generation:

  • ArrivalAllMen_A1-C6 (batch 1, runhex X)
  • ArrivalAllMen_A1-C6 (batch 2, runhex Y)
  • ArrivalAllMen_A1-C6 (batch 3, runhex Z)

(Same for Ladies and Singles.)

This gives 54 images per group (3 batches × 18) instead of 18, dramatically improving the chance of getting keepers with correct count and good composition.


Concurrency Settings (Validated 2026-03-24)

Pro model (grok-imagine-image-pro):

  • Rate limited (~30 images/min)
  • Use serial or low concurrency (1-6 workers)
  • Applies to: Undressed, Dressed

Standard model (grok-imagine-image):

  • Can handle 300+ requests/min
  • Use high concurrency (60 workers) for batch stages
  • Applies to: Topless (A-Z + ABERX), Arrival (individual + ArrivalAll)

Topless specifically: Use 60 workers for A-Z + ABERX batches (dramatically reduces time from ~45min serial to ~5min concurrent).


Identity Handling (MANDATORY)

Critical Rule: Use exact identities from source scripts. Zero improvisation.

Source of truth (in priority order):

  1. Generation script for the tier (render_courtesans_doublepairs_fullynaked.py)
  2. Tier identity reference file ({Tier}/Notes/courtesan_identities.json)
  3. Generation manifest files (as fallback only)

Never:

  • Infer identity from names ("Snowfield" ≠ ice-blue, "Redmane" ≠ specific red shade)
  • Fill in missing details with generic templates
  • Assume identity details "don't matter that much"

Every detail matters:

  • Hair style (braid/bun/loose/half-up) = recognition anchor
  • Face shape (narrow/heart/oval/triangular) = structural identity
  • Eye exact color (ice-blue ≠ blue-gray ≠ pale blue-gray)
  • Distinctive marks (freckles/scars/beauty marks) = verification points

Validated: Swapping Cillian's identity into Eira's proved identity controls WHO renders, not WHERE. Use exact text or get the wrong person.


Hair Color Aging Constraint (All NPCs)

No Gray Hair Under Age 40

Rule: Avoid gray, silver, or salt-and-pepper hair on NPCs under 40 years old.

Why: Gray hair overrides age numbers — models render visibly older (40s-50s) even when prompt says "25-year-old."

Applies to:

  • All genders (males and females both affected)
  • All NPC types (courtesans, merchants, guards, etc.)

Safe hair colors for young NPCs (under 40):

  • Copper-red, auburn, chestnut
  • Ash-blond, platinum-blond, golden-brown
  • Brown (light/medium/dark/deep)
  • Black, dark auburn

Reserve for mature NPCs (40+):

  • Gray, silver, salt-and-pepper
  • "Distinguished silver" (intentional aging)

Note: "Pale blue-gray" or "smoky-gray" eyes are fine; constraint applies to HAIR only.

Discovered: 2026-03-25 — Comfortable tier males rendered old with gray hair; Liv (female, gray hair) consistently looked 40s despite stated age in 20s.