ComfyUI-Ethanfel-Prompt-Builder/docs/krea2-prompt-guide.md

# Krea2 Prompt Guide

This document records prompt rules discovered from actual SxCP generator
outputs tested in Krea2. It is not a generic prompt cookbook. Add a rule only
when an A/B image comparison shows that the wording improves or breaks Krea2
behavior.

## Core Rule

Krea2 responds best when the prompt gives one clear visual hierarchy:

1. subject/cast descriptor,
2. action or pose,
3. clothing state,
4. location,
5. camera/layout,
6. expression,
7. composition/crop,
8. style.

Avoid letting two sections describe incompatible camera or framing intents.

## Prompt Output Contract

- `sxcp_eval_out` must contain only the prompt being tested.
- Analysis, scoring, and generator notes belong in chat or `sxcp_eval_log`.
- Keep one experiment variable per cycle when possible.
- Lock sampler seed, character, location, and camera when testing wording changes.
- Treat the MCP seed as transport metadata. Preserve it for prompt-only A/B tests
  and do not write it into the visible prompt text.

## Positive-Only Conditioning

Never put negative-conditioning phrases inside the positive Krea2 prompt. A
positive prompt should describe only the visual elements that should appear.
Do not write phrases such as `no shaft`, `no hands`, `without clothing`, or
`avoid X` into `sxcp_eval_out`; for distilled Krea2, those phrases can reinforce
or hallucinate the unwanted object.

This SxCP loop has no active negative-output contract. A same-positive,
same-seed probe on seed `424242` compared empty negative conditioning against
strong negative text targeting visible prompt attributes, and the rendered
images stayed visually unchanged. For this workflow, keep Krea2 pose fixes in
positive wording only.

## Seed-Controlled A/B Tests

Use one fixed sampler seed when deciding whether prompt wording helped Krea2.
If the SxCP generator/control seed differs from the sampler seed, record it as
`generator_seed` in the eval log. A single image can justify a prompt-only retry
when the mismatch is obvious, but a generator rule needs either repeated
evidence or a generated prompt that is structurally wrong before rendering.

When a workflow batches soft/hard prompts through an index switch, sidecar text
files may not be the exact prompt used for each rendered image. If the sidecar
and image disagree, inspect the PNG workflow metadata and the final text encode
input before patching the generator.

When reviewing an eval payload, log:

- emitted sampler seed,
- generator/control seed when it differs,
- original generated prompt,
- edited prompt,
- image failure or improvement,
- whether the change should stay prompt-only or become a generator patch.

## Evidence Promotion Threshold

Do not promote a single-character, single-location prompt hack into a durable
guide rule. A prompt-only fix can be recorded as accepted when it cleanly
improves the fixed-seed image, but a general wording rule needs either repeated
evidence across distinct subjects/locations or a source prompt that is
structurally wrong before rendering.

Scene repair words must match the selected location. Do not use a bedroom,
bedding, couch, shower, car, or other scene anchor to solve a foreground problem
inside an office/coworking setup unless that anchor is already present in the
selected scene. If the model needs a foreground filler to avoid hallucinating
extra anatomy, choose an anchor from the actual location or keep the result as
an inconclusive prompt-only retry.

## Camera And Composition

### Orbit / Multiangle Camera

When Krea2 receives an orbit or multiangle camera, avoid selfie-specific wording
unless the intended camera is actually a handheld or mirror selfie.

Works better:

- `lifestyle portrait frame`
- `creator portrait frame`
- `outfit-check pose`
- `wide environmental coworking camera layout`
- `camera placed several meters away`
- `full seated body from head to knees`
- `room depth surrounding the subject`

Conflicting wording:

- `selfie frame`
- `phone selfie`
- `holding the phone`
- `creator-shot phone photo`
- `handheld camera realism`

Observed result: selfie words pulled a back-right elevated wide shot into an
arm-length selfie. Removing selfie terms made the image follow the rear-quarter
view much better.

### Wide Shots

Krea2 tends to keep attractive subjects large in frame. To get a real wide or
environmental frame, be explicit about distance and visible environment.

Useful phrasing:

- `camera placed several meters away across the desk aisle`
- `full seated body from head to knees remains visible`
- `nearby desk edge, laptop corner, repeated desk rows, and tall-window depth clearly readable`
- `wide environmental room framing`

Avoid relying on `wide shot` alone.

## Location Layout

Location-aware camera text works when it describes the room around the subject
without stealing the foreground from the subject.

For coworking lounge:

- Keep `warm desks`, `laptop tables`, `glass partition seams`, `repeated desk rows`,
  `plants`, and `tall windows`.
- Mention foreground anchors only when the camera should actually see them.
- In POV, keep location anchors beside or behind the bodies, not in the lower
  foreground.

## Clothing Continuity

When a softcore outfit is reused in a later branch, name what happens to actual
outfit pieces instead of using generic fabric language.

Works better:

- `denim shorts are pulled aside or removed below the hips`
- `button-down shirt tied at the waist and fitted bralette remain visible from the same outfit`

Avoid generic fallback wording:

- `fabric slipping off`
- `partly exposed`
- `outfit pushed aside where needed`

Use generic wording only when no source outfit exists.

## Climax / Ejaculation Wording

Krea2 needs the visible fluid action stated directly in climax routes. Keep
`semen` in the action sentence when the generator intends an ejaculation or
post-ejaculation image.

Works better:

- `the man ejaculates semen across her body`
- `visible semen lands across her body`
- `ejaculates semen across her lower back, ass, and thighs`

Avoid relying only on softer aftermath wording:

- `visible ejaculation`
- `visible orgasm aftermath`
- `cum on hands and body`

Seed `42052` exposed this as a structural formatter issue: the route selected a
standing/kneeling climax pose, but the Krea prompt dropped `semen` and kept only
generic aftermath language. The generator now preserves direct `ejaculates
semen` wording for this path before the prompt reaches Krea2.

### Ready / Post-Ejaculation Open-Thigh Display

The ready atlas folder is not a neutral setup pose. Treat it as a still
post-ejaculation open-thigh display where the fluid-covered exposed opening is
the first visual priority and the active lower-foreground action is secondary.

Works better:

- `POV post-ejaculation open-thigh display`
- `the wet aftermath detail is the exact center`
- `thick semen and clear fluid cover the exposed pussy and inner thighs`
- `her body stays still after ejaculation`
- `her face and torso remain visible behind the open-thigh frame`
- `office chair seat and chair arms frame the lower foreground`
- `desk edges, laptop tables, glass partitions, plants, and tall-window depth`

Failure modes:

- generic `ready` or setup-before-sex wording
- active penetration or thrusting becoming the main read
- fluid detail present but secondary to a lower-foreground body cue
- bedroom/bedding anchors inserted into coworking scenes
- closed thighs or cropped torso hiding the atlas open-thigh display

Fixed-seed finding: CodexMCPTest source subject `52` with sampler seed
`1123581321` produced first ready/aftermath evidence. The generic baseline kept
the coworking scene and open-thigh pose, but collapsed into active
lower-foreground penetration with fluid secondary. The accepted candidate made
the wet aftermath detail the primary visual event while preserving the
office-chair/coworking setting. Under the category-exit rule this is mirrored
into the generator as a provisional improvement over baseline, while the catalog
route remains candidate pending broader repetition.

### Spread / Open-Thigh Presentation

The spread atlas folder is a direct camera-facing presentation pose. The key
read is the leg geometry first: raised knees or wide thighs form a V-frame
toward the viewer, hands hold the knees or thighs, and the face and torso remain
visible behind the open-thigh frame.

Works better:

- `knees raised and held wide`
- `thighs form a broad V-frame around the centered exposed vulva`
- `hands hold her knees and upper thighs`
- `face and torso remain visible behind the open-thigh frame`
- `office chair seat and chair arms frame the lower foreground`
- `desk edges, laptop tables, glass partitions, plants, and tall-window depth`

Failure modes:

- low seated spread with legs open but no raised-knee V-frame
- missing hands, or hands not anchoring the knees/thighs
- outfit remaining closed over the center of the pose
- partner/contact wording competing with the camera-facing presentation
- scene repair words that turn the coworking office into a bedroom setup

Fixed-seed finding: CodexMCPTest source subjects `50` and `47` with sampler seed
`3141592653` showed the atlas hierarchy improved over generic spread wording on
two visible subjects. Source `50` already produced a coherent low seated spread,
but the candidate added hands on raised knees and a stronger V-frame. Source
`47` failed baseline exposure because the shorts stayed closed; the candidate
opened the center and added raised knees and hand anchors. The generator now
carries this as a provisional improvement over baseline, while the catalog route
remains candidate pending a wider V-frame repeat on another seed.

## POV Outercourse

### Ballsucking / Testicle Sucking

The atlas examples are low first-person pelvis views. The visible partner should
read as bent forward between the viewer's open thighs, with the chest low over
the viewer's pelvis and the face placed below the shaft at testicle height.
Starting from the viewer or using only generic kneeling language can make the
head float too high.

Works better:

- `the woman bends forward and kneels very low between the viewer's open thighs`
- `chest low over the viewer's pelvis`
- `face is below the viewer's penis at testicle height`
- `head below the shaft base, chin near the viewer's pelvis`
- `mouth lower than the shaft base and aligned to the underside/scrotum`
- `balls fill the exact center foreground as the largest contact object`
- `lips and tongue press to the scrotum and balls`
- `shaft is cropped to the side as a landmark, away from her lips`
- `composition priority is balls first, scrotum contact second, woman's low face third`

Partial side-low axis:

- `low side-pelvis POV`
- `face is the closest visible partner part`
- `cheek against the viewer's inner thigh`
- `scrotum and testicles centered against her tongue and lips`
- `balls fill the lower center foreground at her mouth`
- `viewer abdomen and inner thighs close around her face`

Avoid:

- `head tucked under the penis shaft` without testicle-height wording
- generic `kneels in front of him`
- making the viewer the main subject before the visible woman is established
- mid-height head placement
- making the raised shaft the centered visual target above her mouth
- repeating shaft/hand-on-shaft wording before the scrotum/testicle contact is
  visually established
- `hanging` or `pressed from above` wording when it creates suspended-object
  artifacts
- `two testicles against tongue` if it turns into wrong rounded anatomy or a
  different body layout

Fixed-seed finding: sampler seed `238365845574312` showed that low-head wording
improves body placement, but Krea2 can still collapse into generic shaft-focused
oral contact when the raised shaft stays centered above the mouth. The accepted
prompt made the scrotum/balls the largest exact-center object and moved the
shaft to a cropped side landmark. Keep this as a prompt-guide rule, but do not
promote the catalog route to proven until the rule repeats on another controlled
seed.

Expansion weak case: sampler seed `1212121212` repeated across two women did
not validate the side-cropped scrotum rule. The candidates made the balls and
scrotum larger, but the head stayed too high and the mouth still targeted the
shaft/glans. Literal wording like `sucking the testicles with her mouth` helped
one result move the shaft off-axis, but it still used tongue-to-base contact
with the head too high. Next tests should avoid label-like
`testicle-sucking position` wording and make head height the primary axis:
head below the shaft base, chin near the viewer pelvis, mouth lower than the
balls/scrotum contact point.

Threshold search update: sampler seed `1212121212` then received about fifty
positive-only prompt probes on one subject plus six repeats on a second subject.
The under-base, chin-near-pelvis, hand-held-side, and direct
`balls in mouth` wording families mostly collapsed back to generic shaft/glans
oral contact. The repeatable partial axis was the side-low family:
`low side-pelvis POV`, `cheek against the viewer's inner thigh`, and
`scrotum/testicles centered against her tongue and lips`. Turns `149`, `162`,
`167`, and `170` show that this can move the contact toward the balls/scrotum,
and tongue-on-testicles is an accepted improvement over the baseline shaft/glans
collapse. Mirror this side-low partial axis into the generator provisionally,
but keep the route queued for broader seed/source expansion because it still
does not reliably produce lips fully wrapped around the testicles.

Generated-route validation: turn `172` showed that the first generator patch
could keep tongue/testicle contact low but still let the woman's torso dominate
the composition. Adding `face is the closest visible partner part` produced
turn `173`, where the face stayed low beside the viewer's thigh and the
tongue/testicle contact remained centered. Keep the face-priority cue in the
generator wording and continue testing because body/rear prominence can still
steal the pose.

Fresh-seed side-camera weak case: sampler seed `5757575757` repeated the
generated side-low route and three contact-priority variants across two women.
Turns `217`-`224` looked strong as side-camera oral views: the face stayed low
between the viewer abdomen/thigh foreground, coworking depth held, and the
cheek/inner-thigh relationship stayed readable. They did not improve exact
ballsucking contact because the mouth and tongue still targeted the shaft/glans.
Record this as useful side-view oral wording, not proof that text alone controls
the testicle-contact object. Bookmark:
`docs/krea2-visual-bookmarks/ballsucking-side-camera-turn221-side-low-oral-weakcase.png`.

Fresh-seed partial update: sampler seed `6262626262` retested the route across
two visible women with twelve positive-only probes. The generated-route
controls on turns `250` and `256` preserved the low coworking side-pelvis view
but still drifted toward shaft/glans oral contact. The best repeatable branch
was turns `252` and `258`:

- `scrotum is the mouth surface`
- `testicles resting across her open lips`
- `tongue cups them from below`
- `head lies sideways on the viewer's inner thigh`
- `chin close to the viewer pelvis`

This is still a partial, not a proven full mouth-wrapped testicle-sucking
result. It is better than the generated-route controls because the contact
moves toward the scrotum/testicles while retaining the atlas side-low
cheek/thigh relationship. Turns `253` and `259` showed a useful secondary
tongue-cradle underside branch, but it reads more centered/frontal than the
side-low atlas family. Patch the provisional generator route to the open-lips
scrotum-surface hierarchy and keep the route queued for more seeds.

Generated-route validation: turns `262` and `263` rendered the patched route
wording directly after the generator edit. Both validation renders kept the low
side face/thigh relationship and moved the tongue/lip contact to the underside
scrotum/testicle area with less shaft-tip dominance than the generated-route
controls on turns `250` and `256`. Keep this as a provisional improvement, not
a proven route, because the result remains tongue/lips-on-testicles rather than
fully mouth-wrapped testicle sucking.

Fresh-seed target-object repeat: sampler seed `9797979797` retested the current
open-lips route and four target-object branches across two visible women. The
generated-route control could succeed on turn `286`, but the blonde control on
turn `291` still drifted toward shaft/base contact. The best repeatable branch
was `scrotum_skin_to_lips` on turns `288` and `293`:

- `scrotal skin is the nearest mouth surface`
- `both testicles rest against her tongue from below`
- `cheek and jaw compressed into the thigh`
- `chin tucked close to the pelvis`

Treat this as a narrow target-object refinement, not a full solution. It
improves the partial hierarchy over the weak control while preserving side-low
cheek/thigh geometry, but it still reads as tongue/lips-on-testicles rather
than fully mouth-wrapped testicle sucking. The `balls_first_center_object`
branch is unsafe as default because turn `290` overfit into a multi-subject or
body-layout artifact.

Generated-route validation: sampler seed `9898989898` rendered the patched
scrotal-skin route directly across two visible women. turns `296` and `297`
both preserved side-low cheek/thigh geometry, kept the face low between viewer
abdomen and inner-thigh cues, and placed scrotum/testicles at the tongue/lip
contact surface. This validates the route patch as a target-object improvement,
but it remains partial because the result is still tongue/lips-on-testicles,
not fully mouth-wrapped testicle sucking.

Fresh-seed weak case: sampler seed `5959595959` retested the current generated
route and three stronger mouth-contact axes across three visible women. The
generated controls on turns `307`, `311`, and `315` kept useful low-pelvis
framing but still drifted toward shaft/glans contact. The `lip-oval` branch on
turns `308`, `312`, and `316` made the scene more upright and shaft-centered.
The `sideways mouth pocket` branch on turns `309`, `313`, and `317` preserved
the best cheek/thigh side-low geometry, but the mouth still targeted the shaft
or base rather than wrapping the testicles. The `chin-pelvis upward seal`
branch on turns `310`, `314`, and `318` became a frontal shaft-focused oral
read. Do not patch the generator from these branches; record this as another
target-object weak case and continue exploring stronger control or a different
positive wording axis.

Fresh-seed occlusion weak case: sampler seed `6060606060` tested a different
target-control strategy after the mouth-shape words failed. The generated
controls on turns `319`, `323`, and `327` were still the best partials,
preserving side-low geometry and some visible scrotum/testicle contact. The
`scrotum foreground occlusion` branch on turns `320`, `324`, and `328` became
upright shaft-centered oral instead of making the scrotum cover the mouth. The
`under-scrotum tongue shelf` branch on turns `321`, `325`, and `329` kept some
low-pelvis framing but still targeted the shaft/glans. The `hand-guided scrotum`
branch on turns `322`, `326`, and `330` made hands and shaft ownership
dominate. Do not patch these axes into the generator; future attempts should
change the target-object strategy or use stronger control rather than adding
more occlusion or hand-support synonyms.

Fresh-seed mouth-axis mixed case: sampler seed `6161616161` tested five new
positive-only axes across three visible women after the occlusion/hand-support
batch failed. The generated-route controls on turns `331` and `337` gave the
best repeated partials: low pelvis framing, face close to the inner thigh, and
tongue/testicle contact. The third generated control on turn `343` still
collapsed onto shaft contact, so the route remains candidate.

Failed or non-repeatable axes:

- `exact mouth-sucking`: literal `sucking the testicles with her mouth` wording
  still became shaft/glans-focused on the clearest inspected branch.
- `single-testicle`: putting a single testicle between the lips did not make the
  model hold the testicle as the mouth target.
- `hanging balls below shaft`: separating the target below the lifted shaft
  mostly preserved shaft/glans oral.
- `side mouth wrap`: turn `335` produced a useful side-oriented partial, but
  turns `341` and `347` collapsed back to shaft contact.
- `chin-pelvis lower-mouth target`: turn `348` produced a useful frontal
  target-object partial with the mouth lower and testicles visible at the
  tongue/lip surface, but the branch is not repeatable enough for generator
  wording.

Conclusion: keep the current provisional generated route as the best repeated
partial for now. Do not patch exact-mouth, single-testicle, hanging-balls,
side-mouth-wrap, or chin-pelvis lower-mouth wording into the generator from
this batch. The full mouth-wrapped target still needs broader expansion or
stronger control.

Fresh-seed pelvis-valley weak case: sampler seed `7171717171` tested
body-plane and camera-height wording after mouth-shape and target-object words
kept failing. The useful atlas comparison was that refs such as `101` place the
viewer flatter and lower, with the partner tucked into the pelvis valley
between the thigh roots.

The `flat pelvis-valley` branch repeated the strongest spatial improvement on
turns `350`, `356`, and `362`:

- viewer reads flatter and more supine;
- inner thighs become side walls around her face;
- her face is tucked lower inside the thigh-wall pelvis valley;
- coworking depth remains location-compatible.

However, this was a body-plane correction, not a target-object correction. All
three repeated flat-pelvis results still made the mouth target the shaft, so do
not patch `flat pelvis-valley` into the ballsucking generator route as-is. Treat
it as a useful orientation hint for stronger control or for a separate
flat-pelvis oral branch.

Other branches from this batch were weaker:

- `thigh tunnel` and pubic-hair mouth-line wording mostly repeated the same
  shaft-centered target.
- `low-cushion chin anchor` drifted into wrong open-thigh/presentation geometry.
- `pelvis-edge target-first` also drifted toward open-thigh/presentation layout
  or target ambiguity.

Conclusion: body-plane wording can fix the atlas orientation while making the
target object worse. Score these separately. For this route, the current
side-low scrotal-skin generated wording remains the better ballsucking partial.

Fresh-seed flat-target hybrid weak case: sampler seed `7272727272` combined the
best prior body-plane wording with the best prior target-object wording across
three women. The tested branches were `flat-valley scrotal-skin target`,
`valley-floor open lips`, `upper-frame shaft lower-scrotum`,
`cropped upper-shaft valley-mouth`, and `side-low flat-valley hybrid`.

The hybrid did not beat the current generated route. The
`flat-valley scrotal-skin target` branch repeated the flatter body-plane read
on turns `368`, `374`, and `380`, but all three stayed shaft-centered. The
`side-low flat-valley hybrid` branch kept some useful side-low look hints on
turns `372`, `378`, and `384`, but the target contact remained ambiguous or
shaft-centered. The upper-frame and cropped-shaft branches also reinforced the
wrong target.

Conclusion: stop the ballsucking text-only prompt loop here for now. Keep the
current side-low scrotal-skin partial as the provisional generator route, keep
the pose queued as candidate/weak, and treat the full target as likely needing
stronger control rather than more positive-prompt synonyms.

### Footjob

The atlas examples are frontal first-person views. The viewer reclines with
thighs framing the lower foreground, the woman sits opposite with her body and
face behind her feet, and the feet are the main contact object around the
upright shaft.

Works better:

- `tight frontal POV footjob close-up`
- `large overlapping soles dominate the lower center foreground`
- `upright shaft compressed between the two soles`
- `a narrow vertical strip of shaft and glans rises between her pressed soles`
- `inner arches and toe pads press inward against the shaft`
- `toes curled around both edges of the shaft`
- `her face, torso, and open thighs remain visible behind her large foreground feet`

Alternate branch:

- `one bare foot crosses horizontally in front of the shaft`
- `sole presses sideways into the glans and upper shaft`
- `other foot braces from the opposite side near the base`

Avoid:

- generic `uses her feet on the viewer's penis` without overlap/contact pressure
- `soles on either side` if it leaves the shaft standing free
- feet too small or too low at the base of the shaft
- hands or mouth competing with the feet as the active contact

Fixed-seed finding: sampler seed `238365845574312` showed that generic footjob
wording can produce the correct frontal POV layout, but the feet may sit beside
the shaft rather than visibly compressing it. The accepted prompt used large
overlapping soles and a shaft trapped between the two soles, with only a narrow
visible strip between the feet. Keep this as a prompt-guide rule, but do not
promote the catalog route to proven until it repeats on another controlled seed.

Expansion finding: sampler seed `3434343434` repeated across two women showed
the baseline was already a valid frontal POV footjob, but `large two-sole clamp`
wording improved foot scale and contact compression on turns `176` and `180`.
The cross-foot side-press wording produced atlas-ref-`86` style alternates on
turns `177` and `181`, but it is a separate branch and weaker as the default.
Mirror the two-sole clamp wording into the generator provisionally and keep
cross-foot wording as an alternate guide branch.

Generated-route correction: turn `183` showed that `two large soles dominate`
is not enough by itself; the route can turn into feet presented to the viewer
and hide the actual contact. Turns `184`, `185`, and `186` restored the target
when the prompt said the soles `clamp the upright shaft between them` and `the
glans rises between the compressed feet`. Turn `188` validated the patched
generated route with the same cue: large soles stayed dominant while the shaft
and contact remained readable. Keep that visibility cue in generator wording
and treat cross-foot side-press as a separate alternate branch.

Fresh-seed route tightening: sampler seed `6868686868` repeated the generated
route and three contact variants across two women. Turns `225` and `229` showed
that the generated route can still hide the shaft/contact behind large soles.
Turns `227` and `231` repeated the stronger default: `large overlapping soles`
with `a narrow visible strip of shaft and glans` rising between the compressed
feet. Turn `226` supported the same center-gap visibility axis; turns `228` and
`232` kept cross-foot side-press useful as a separate alternate. Turn `233`
validated the patched generated route with overlapping soles, the centered
shaft/glans strip, viewer body cues, and coworking depth intact. The generator
route now uses the overlapping-sole/narrow-strip wording.

Fresh-seed promotion: sampler seed `7373737373` repeated the patched generated
route across two visible women. Turns `264` and `267` used the exact
overlapping-sole/narrow-strip generated route and preserved large foreground
sole dominance, visible centered shaft/glans strip, toe/arch pressure, face and
torso behind the feet, and coworking depth. Turns `265` and `268` repeated the
tight center-gap visible-glans branch as supporting evidence. Turns `266` and
`269` remained valid cross-foot side-press alternates, but they are not the
default because they give up the two-sole enclosure. Treat the default frontal
footjob route as proven; keep cross-foot side-press as an alternate guide
branch.

### Fingering / Manual Stimulation

The atlas examples are close first-person manual-contact views. The visible
partner reclines with open thighs toward the camera, one foreground hand enters
from the lower frame, and the hand is the main contact object. Generic
male-participant POV wording can make Krea2 insert lower-foreground anatomy that
competes with the hand.

Works better:

- `single foreground hand is the largest lower-frame object`
- `wrist entering from bottom center`
- `two fingers placed at her vulva and clit as the clear contact point`
- `open thighs forming a V around one camera-side hand`
- `manual contact first, open-thigh geometry second, woman face and torso third`
- `office-chair support and coworking depth fourth`

For coworking or office scenes, keep the foreground and depth anchors from the
selected location:

- `black office chair seat and chair arms frame the lower foreground`
- `near desk edge`
- `laptop table corners`
- `glass partition seams`
- `repeated desk rows`
- `plants and tall-window depth`

Failure modes:

- generic `viewer fingers her` wording without hand scale or ownership
- male-participant POV language that makes the lower foreground the main subject
- bedding or soft-surface anchors inside an office/coworking scene
- multiple foreground body cues competing with the active hand

Fixed-seed finding: sampler seed `238365845574312` repeated across
CodexMCPTest source subjects `46` and `47`. In both baselines, generic
fingering/manual-contact wording collapsed into shaft-forward lower-foreground
anatomy. The accepted prompt kept the coworking scene intact and changed only
the hierarchy: one large foreground hand, open thighs, woman face/torso, then
office-chair/coworking depth. Keep this as a prompt-guide rule, but do not
promote the catalog route to proven until it repeats on another seed or location
family.

Matrix update: CodexMCPTest source subject `50` with sampler seed `987654321`
repeated the same pattern. The generic baseline kept the coworking setting but
collapsed into shaft/body-dominant lower-foreground anatomy; the hand-first
candidate preserved the coworking office setting and produced readable active
fingers. This strengthens the prompt-guide rule and adds the required second
sampler seed. The manual/fingering generator now carries this hierarchy as a
provisional improvement over baseline, while the catalog route remains
`candidate` until broader evidence proves it.

Weak-case update: CodexMCPTest source subject `52` with sampler seed
`1357913579` shows the current rule is still too early for generator mirroring.
The generic baseline failed in the same lower-foreground way, and a camera-side
hand-first candidate still kept the wrong lower-foreground anatomy dominant. A
second candidate succeeded only after making the active hand explicitly the
woman's own right hand and using the office chair cushion/arms as the positive
lower-frame structure. Keep this as a prompt note, not a generator rule for the
participant-hand POV route, because the successful candidate changed hand
ownership semantics.

Experimental generator-mirror probe: generator seed `4303` with sampler seed
`238365845574312` showed that a generated-route prompt using the hand-first
hierarchy can preserve the office/coworking setting and remove the shaft-forward
foreground artifact from generic baselines. This is not enough to justify a
production generator patch. Collect a broader seed, subject, and location matrix
before changing route code for this case.

### Wand Toy Contact

The atlas examples are close first-person toy-contact views. The visible
partner reclines or sits back with open thighs toward the camera, face and torso
behind the leg frame, and one foreground hand holds a wand-style massager with
the rounded head at the central contact point.

Works better:

- `single continuous teal wand-style massager is the largest lower-frame object`
- `rounded bulb head presses flat to her vulva and clit as the central contact point`
- `smooth handle angles in from the bottom right inside the viewer's visible hand`
- `open thighs and knees form a V around the foreground wand`
- `wand head contact first, open-thigh geometry second, visible hand and handle third`
- `woman face and torso fourth, coworking depth fifth`

For coworking or office scenes, reuse the location anchors instead of adding
bedroom surface cues:

- `office chair seat and chair arms frame the lower foreground`
- `desk edges`
- `laptop tables`
- `glass partitions`
- `plants and tall-window depth`

Failure modes:

- generic `handheld wand vibrator is used near her` without device continuity
- Krea2 splitting the cue into a contact toy plus a second wand-like foreground object
- toy floating without a visible hand or handle
- the woman holding the toy when the intended view is a foreground viewer hand
- mouth, foot, penetration, or manual-contact wording competing with the wand

Fixed-seed finding: CodexMCPTest source subject `48` with sampler seed
`246813579` produced first wand evidence. The generic baseline preserved the
coworking setting and found a toy-contact pose, but split the prompt into two
wand-like objects. The accepted candidate made one continuous teal wand device
the primary lower-frame object, with a readable handle grip and bulb-head
contact. Treat this as first prompt-guide evidence only; repeat it on another
source or seed before changing generator defaults.

Expansion finding: sampler seed `8642086420` repeated the teal lower-right
single-device axis across two visible women on turns `190` and `194`, improving
over generic baselines `189` and `193`. The black close-bulb wording drifted
toward face/mouth contact, while the upper-left pale wand is useful as an
atlas-`107` alternate but weaker as a default. Turn `197` validated the patched
generated route from generator row seed `46`: one continuous teal wand, visible
viewer hand, central bulb contact, and coworking office-chair depth all
survived. Mirror this axis into the generator provisionally and keep the route
queued for another source/seed expansion.

Proven-route repeat: sampler seed `7979797979` repeated the generated teal
lower-right route across two women on turns `234` and `238`. Device continuity,
visible viewer hand/handle, central bulb-head contact, open-thigh chair
geometry, visible subject, and coworking depth all survived. Contact-first teal
variants on `235` and `239` also worked. The white upper-left branch repeated
on `236` and `240` as a valid alternate for atlas refs with diagonal handle
entry. The large-bulb branch on `237` and `241` is rejected as the default
because it can hide exact contact and overtake the lower frame. The catalog
route is now proven with the teal lower-right single-continuous-device default.

### Boobjob / Titjob

The atlas examples are frontal and upright: the visible partner faces the viewer,
kneels between the viewer's thighs, and compresses the shaft between the breasts.
Forward-bent wording can still place the body correctly, but it weakens the
breast contact.

Works better:

- `POV boobjob position`
- `woman kneels upright between his legs facing him`
- `penis rises vertically in the lower foreground`
- `squeezed between her pressed-together breasts`
- `woman's own fingers and nails cup her breasts from the outside`
- `glans emerging above the cleavage directly below her mouth`

Avoid vague or conflicting wording:

- `torso bent forward over his pelvis`
- `both hands push her breasts` without naming whose hands
- `only foreground hands` when the intended hands are the woman's hands

### Handjob

The atlas examples are direct and simple: the woman faces the viewer between his
thighs, her hand is the visible contact point, and her face or torso stays behind
the shaft. In fixed-seed tests, naming `one hand` plus a generic POV camera line
could make the lower hand read as the viewer's hand. Naming the woman's hands
made the action clearer.

Works better:

- `POV handjob position`
- `woman kneels between his legs facing him`
- `the woman's right hand wraps around the viewer's penis`
- `her left hand steadies the base`
- `viewer thighs and pelvis frame the lower edges`
- `without his hands covering the action`

Avoid:

- generic `one hand grips` when hand ownership matters
- allowing `foreground hands` to compete with the woman's active hand

## POV

In POV prompts, the visible subject should still be established first. The POV
participant is the camera viewpoint, not a normal visible cast member.

Works better:

- visible subject descriptor first,
- then POV action,
- then foreground hands/body/clothing cues.

For POV clothing, describe only visible body/clothing fragments:

- `foreground hands, hips, thighs, or lowered waistband`
- `foreground hands, forearms, sleeves, or torso edge`

Avoid:

- full third-person `Man A wears...` phrasing for the POV participant,
- making `the viewer` the first subject before the visible character is
  established.

For POV climax wording, the fluid target must follow the pose before expression
tokens. Rear-entry, doggy, bent-over, face-down, and on-all-fours poses should
target the ass, thighs, and lower back even if the expression detail mentions
face, lips, mouth, or tongue.

Evidence:

- Dataset seed `52` generated an internally contradictory prompt: on-all-fours
  rear-view positioning paired with a face/chest ejaculation target.
- Corrected seed `52` and follow-up seed `5202` both rendered the rear-view
  target consistently when the wording used `across her ass, thighs, and lower
  back` and kept the clothing state tied to the lower garment.

### Blowjob Top View / Overhead Vertical Shaft

The atlas examples are steep overhead first-person oral views. The useful
structure is not generic kneeling oral and not just a vertical shaft; it is a
nadir-angle standing male POV where the nearby floor plane dominates and the
woman is directly below the viewer between his feet.

Works better:

- `nadir-angle standing male POV top-view oral position`
- `viewer looks almost straight down from his torso toward the floor`
- `nearby carpet/floor plane dominates the image`
- `viewer abdomen, shorts, thighs, and feet frame the lower foreground`
- `shaft appears as a short centered vertical column from the foreground`
- `one woman kneels directly below the viewer between his feet`
- `hair crown, forehead, shoulders, hands, and knees are visible from above`
- `desk legs, chair wheels, carpet texture, and floor seams as top-down office anchors`
- `mouth seals around the centered shaft`
- `one hand wraps the base`
- `face, shoulders, and compact foreshortened torso remain readable beneath the shaft`

Failure modes:

- generic kneeling oral that drifts into side-profile framing
- off-center shaft/mouth alignment
- hand present but not supporting the vertical centerline
- a close face crop that loses the viewer-underbody lower edge
- `plumb-line` or `map` wording that literalizes into drawn graphics

Fixed-seed finding: CodexMCPTest source subjects `46` and `47` with sampler
seed `4242424242` showed that generic blowjob/oral wording can already produce
a usable top-down image, but the atlas hierarchy tightened the shaft-to-mouth
axis. Source `46` gained clearer hand-at-base structure; source `47` moved from
a side-profile/off-center read into a centered vertical top-down read. The
generator now carries this as a provisional improvement over baseline, while
the catalog route remains candidate pending another seed.

Correction probes: a later same-seed source-`46` probe showed that vertical
center-shaft wording and generic steep-overhead wording can still feel too
frontal or horizontal. A short axis loop found the stronger terms: `nadir-angle`
or `bird's-eye` paired with `standing male POV`, `floor plane dominates`, nearby
top-down office anchors, `one woman directly below between his feet`, and
`short centered vertical column`. The `plumb-line` and `map` terms produced
good geometry but literal drawn artifacts, so they should stay out of generator
wording.

### Blowjob Side Profile / Side-Phone Weak Case

Seed `5656565656` showed that current side-profile oral wording is good at a
related composition, but that composition is not valid POV evidence for
`pov_blowjob_side_profile_oral`. It reads like a side-phone or external
side-camera shot taken from beside the man rather than a true first-person
viewpoint from his eyes or body line.

Useful for a future non-POV route:

- `side-phone oral composition`
- `external side-camera oral view`
- `woman kneels or leans beside the man's pelvis`
- `side-facing mouth and cheek profile remain readable`
- `office/coworking depth remains visible beside and behind the bodies`

Visual bookmark: turn `202`
[`blowjob-side-camera-turn202-awesome-side-phone.png`](krea2-visual-bookmarks/blowjob-side-camera-turn202-awesome-side-phone.png)
is a strong side-camera / side-phone composition for a future non-POV route. It
should not be counted as POV evidence, but it is a good look target for the
side-camera family. The original output filename is also preserved as
[`blowjob-side-camera-img_5a776aafda8f4a80ab12d995e4ce8020.png`](krea2-visual-bookmarks/blowjob-side-camera-img_5a776aafda8f4a80ab12d995e4ce8020.png).

Do not count as POV evidence:

- attractive side-camera framing where the camera is beside the man
- the viewer's body appearing as a photographed subject rather than the camera
  owner
- side profile that loses the first-person lower-body viewpoint

Fixed-seed weak case: CodexMCPTest source subjects `46` and `47` with sampler
seed `5656565656` produced four attractive side-phone / external side-camera
renders. The stricter candidate improved profile readability in some ways but
did not make the result more POV. Keep this wording for a later non-POV
side-camera route; do not patch the POV generator from it.

Single-source improvement: the same seed on source `46` improved when the prompt
made the foreground owner explicit as the adult male viewer:

- `visible adult male chest, navel, abdomen hair, pelvis, and thighs form the
  central foreground surface`
- `woman lies on her side beside his left hip`
- `her torso lateral along the left edge of the frame`
- `side-facing mouth seals around the shaft near the center of the male abdomen`
- `one of her hands wraps the base directly below her lips`

This produced an atlas-like first-person body-line view with the male abdomen
and navel as the central foreground and the woman entering laterally from the
side. Treat it as accepted single-source evidence, not a generator rule.

Fragility warning: source `47` on the same sampler seed rejected a related
`body-axis` candidate because Krea2 transferred the body-axis cue to the
woman's torso/body, producing a seated or straddling non-POV composition. Future
side-profile oral tests should name the foreground owner repeatedly as the
adult male viewer and verify that the woman's body stays lateral instead of
becoming the central body surface.

Matrix update: sampler seed `9753197531` repeated the corrected side-profile
body-line axis across two visible women. Generic baselines on turns `198` and
`202` remained attractive side-phone or external side-camera reads. Pure
`male_body_axis` wording on turns `199` and `203` was fragile because it could
make the male read as the photographed subject. The repeatable winner was the
`lateral_edge_entry` axis on turns `200` and `204`: the male viewer's abdomen,
navel, pelvis, and near thigh create a broad horizontal foreground body
surface; the woman enters laterally from the left edge beside his hip; cheek
and jaw stay in profile; mouth contact sits at the male abdomen line; and one
hand stays around the base under her lips. Mirror this lateral-edge body-line
hierarchy into the side-lying oral generator route provisionally, while keeping
the variant queued for broader seed/source expansion.

Generated-route validation: turn `206` proved that the patched generated route
can preserve the male abdomen body-line, lateral side-profile entry, and
coworking depth, but it let the mouth/tongue float above the shaft while the
hand became the contact anchor. Adding `lips touching the shaft at the male
abdomen line` and `mouth-to-shaft contact is the nearest facial detail`
produced turn `207`, which preserved the body-line composition and moved the
oral contact onto the shaft. Keep those contact-priority terms in the generator
wording, and keep the route provisional because the result is still closer to
mouth/tongue contact than fully mouth-wrapped side-profile oral.

Fresh-seed weak case: sampler seed `8484848484` retested the side-profile
body-line route across two visible women. Generated-route turns `270` and
`273` preserved some useful male body-line foreground and mouth contact, but
still read partly as external side-camera compositions. Abdomen-line contact
lock turns `271` and `274` made Krea2 transfer the body-line cue onto the
woman's torso/body, so the woman became the photographed foreground subject.
Lateral-edge mouth-first turns `272` and `275` improved attractive side-camera
mouth contact, but the male appeared as a photographed subject at the frame
edge. Treat this as a weak-case record: do not promote the route and do not
patch from these branches. Future probes need stronger viewpoint ownership
without making either participant a photographed side-camera subject.

Fresh-seed lower-right torso repeat: sampler seed `9595959595` retested the
current generated route against body-owner wording across two visible women.
The current route controls, turns `276` and `280`, preserved contact, but the
blonde control still exposed the male as a photographed side subject. The
`lower_right_torso_anchor` branch repeated on turns `279` and `283`: `adult male viewer's own torso starts at the lower edge`,
runs into the lower-right foreground, and navel, abdomen hair, pelvis, and
near thigh mark the camera owner's body. This improved viewpoint ownership
while keeping lateral side-profile mouth contact. Mirror this lower-right
torso anchor into the side-lying oral generator provisionally, keep the route
queued, and continue rejecting the attractive side-camera branch as non-POV
evidence.

Generated-route validation: sampler seed `9696969696` rendered the patched
side-profile route directly across the same two visible women. Turns `284` and
`285`; turns `284` and `285` both kept the lower-right own-body foreground, navel/abdomen hair/pelvis
and near thigh as camera-owner landmarks, side-profile mouth contact at the
abdomen line, and coworking depth. This validates the provisional route patch
against the previous photographed-male-face side-camera failure, but it is not
enough to promote the pose; keep expanding across source images and seeds.

Fresh-seed promotion: sampler seed `5858585858` retested the patched generated
route across three visible women, with two conservative wording branches beside
each control. Generated-route controls on turns `298`, `301`, and `304`
repeated the same usable structure: the male abdomen/navel/pelvis/near thigh
remain the lower-right camera-owner body plane, the woman enters laterally in
side profile, mouth contact stays at the abdomen line, and the coworking depth
remains behind the bodies. The lower-edge owner branch on turns `299`, `302`,
and `305` strengthened body ownership but did not beat the default cleanly.
The `side-camera-style self-body crop` branch on turns `300`, `303`, and `306`
looked strongest aesthetically, but it sits closer to the non-POV side-camera
family. Promote the current generated side-profile POV hierarchy to proven and
keep the side-camera-style wording as a future look branch rather than the
default POV route.

### Blowjob Laying Frontal / Wide V-Frame

For prone frontal oral, generic `woman lies belly-down between his open thighs`
wording can already produce usable centered oral contact. The fixed-seed
improvement is the atlas hierarchy around the viewer's thigh frame and the
woman's low body line.

Useful wording:

- `open thighs form a wide symmetrical V-frame from the lower corners toward
  the center`
- `pelvis and lower abdomen at the near edge`
- `woman lies belly-down in the gap between his thighs`
- `torso stretched low and horizontal`
- `hips and legs trailing away behind her along the center line`
- `hands wrap the base`
- `centered mouth-to-shaft contact`

Fixed-seed finding: CodexMCPTest source subjects `46` and `50` with sampler
seed `6767676767` improved over already-usable baselines. The baseline prompts
preserved coworking continuity, hands, and centered mouth contact, but tended
to read as raised-hips or all-fours rather than prone belly-down. The candidate
kept the subject/outfit/office scene and strengthened the wide V-frame and
low-horizontal torso hierarchy. Source `46` still kept some high-hip posture,
so this is a provisional generator improvement, not a proven rule.

### Blowjob Sitting Upright / Low Mouth Contact

For upright sitting oral, generic `mouth aligned to the centered shaft` wording
can preserve the seated office layout but leave the face too high, making mouth
contact implied rather than visible. The useful fixed-seed change is to make the
mouth-contact point the central hierarchy while preserving the seated torso.

Useful wording:

- `viewer reclines with open thighs forming the lower V-frame`
- `lower abdomen anchoring the near edge`
- `woman sits low between his open thighs`
- `torso upright behind the action`
- `face lowered close to the exact center contact point`
- `vertical shaft rises from the exact lower center`
- `open mouth covers the tip at the centerline`
- `both hands stay low at the base directly below her mouth`

Fixed-seed finding: CodexMCPTest source subjects `46` and `50` with sampler
seed `7878787878` improved when the candidate replaced high-face alignment with
low-mouth seated hierarchy. Both baselines preserved coworking continuity and
the viewer-thigh frame but kept the woman's face too high and hands beside the
base. A first source-`46` candidate that over-weighted vertical shoulders made
the mouth drift farther away. The accepted wording lowered the face to the
centered tip and made hands-at-base structure readable on both sources. Source
`50` had some outfit looseness/drift, so this is a provisional generator
improvement, not a proven rule.

### Missionary Open-Leg / Seated-Lounge Drift

The first open-leg missionary A/B did not justify a generator patch. It showed
that office-lounge support words can preserve coworking continuity but also
push the result toward a seated or angled cushion route. That route can still be
a valid missionary angle; it is just not the flatter elevated-support atlas
subcase.

Useful partial cues from seed `8989898989`:

- `viewer between her legs from the lower foreground`
- `hands hold the undersides of her thighs`
- `open thighs form a broad V-frame around the centered contact line`
- `lower abdomen and hands anchor the lower foreground`
- `face and torso visible behind the open thighs`

Weak points:

- `low office-lounge sofa surface` and visible sofa backs made both source
  subjects read like an angled lounge-seat missionary route
- the candidate improved source `50` thigh framing and hand placement, but did
  not materially improve source `47` over its already-strong baseline
- neither candidate fully matched the flatter elevated-support references

Follow-up prompt-axis probes on the same sampler seed showed a useful but still
unproven direction. Pose-first prompts with `flat floor mat`, `thin floor
cushion`, `wide V-shaped tunnel`, `diagonal thigh rails`, `viewer hands grip
the thighs from below`, and `coworking carpet, floor seams, desk legs, and
chair wheels stay low around the cushion` improved the flat-back and open-thigh
geometry. However, those pose-first probes omitted the original subject/look
block, so they only count as geometry-axis discovery. Subject-look-first
follow-ups preserved the woman look better, especially with the `V-shaped
tunnel` wording, but still stayed in the angled cushion family. Do not treat
that as invalid missionary; treat it as the wrong subcase when the target is the
flat elevated-support atlas view.

A subject-first follow-up batch on turns `77`-`80` confirmed the same weak
case. Keeping the original woman/look block first preserved identity and face
adherence across all four candidates, but stronger phrases like `camera mounted
just above the male pelvis`, `head, shoulders, spine, hips, and hair all
resting on the same low horizontal plane`, `wide thigh gate`, `overhead
first-person missionary`, and `thin cushion placed directly on a coworking
carpet` still did not beat turn `76`. The best candidate, turn `80`, kept
coworking depth, viewer hands, and the V-frame readable, but still looked
seated/upright on a cushion rather than the flat elevated-support atlas
subcase.

The successful split came from atlas refs where the viewer is not reclined with
thighs projecting forward. Turns `81`-`84` tested elevated supports and viewer
edge placement. The accepted axis is:

- `flat elevated support` or `rounded white lounge platform`
- `standing or braced man at the foot edge`
- `feet and vertical shins visible below the support edge`
- `legs drop down along the side edges below the support`
- `woman lies flat across the tabletop/platform`
- `hands holding her calves or outer thighs at the edge`

Turn `84` was the best subject-first candidate: the woman stayed flat on one
rounded platform plane, office depth remained coherent, viewer hands held the
outer thighs, and the lower edge read as a support edge instead of a reclined
viewer-thigh foreground. Turn `81` was useful secondary evidence with a white
table/support and feet below the edge. Turn `82` is a weak branch: `kneeling
tightly at the foot edge` hallucinated an extra male body behind her, so do not
make kneeling-edge wording the primary cue. Patch only the raised-edge or
edge-supported route with this elevated-support wording; keep generic
missionary available for valid angled missionary views.

### Missionary Folded / Contact-First Knee Block

Folded/high-leg missionary can be reached with subject-first wording, but Krea2
will often let the knees and feet become the whole image. The first fixed-seed
batch on seed `8989898989` showed that `compact knee block`, `vertical thigh
columns`, `ankles and feet close to the camera`, and `hands holding calves`
produce folded geometry, but the visible shaft/contact can disappear.

The accepted follow-up axis from turn `89` is contact-first:

- `viewer lower abdomen anchors the bottom edge`
- `large centered shaft rising from the lower center`
- `shaft/contact enters the exposed vulva below the folded knees`
- `both knees folded tightly toward her chest into one compact knee block`
- `viewer hands hold her calves`
- `face, breasts, and torso remain visible behind the raised knees`

Turn `89` was the strongest combined result: it preserved the same subject/look
and coworking platform, recovered a visible lower-center shaft/contact, and
kept the knees folded high as a compact block. Turn `92` was useful secondary
evidence for pelvis/shaft anchoring, but the fold was looser. Prefer the
turn-`89` hierarchy over abstract `penetration line` wording. Mirror it only
into the folded-missionary route as a provisional patch; keep generic
missionary and elevated-edge missionary separate.

### Cowgirl Frontal / Wide Thigh Bridge

Frontal cowgirl is not currently a structural generator failure. On seed
`8989898989`, the generic subject-first baseline already produced a valid
woman-on-top POV: the viewer was underneath, the woman faced him, torso stayed
upright, coworking depth remained coherent, and lower-center contact was
readable.

The useful atlas refinement is the wide-thigh silhouette. For refs like
`5.cowgirl/100_cowgirl.png` and `5.cowgirl/101_cowgirl.png`, prefer:

- `viewer reclines underneath her`
- `viewer lower abdomen and pelvis at the bottom edge`
- `wide horizontal thigh bridge from left edge to right edge`
- `knees planted outside the viewer's hips`
- `torso upright above the centered contact point`
- `centered shaft contact below her belly`
- `viewer hands gripping the sides of her thighs`

Turn `95` was the strongest same-seed candidate because it made the thighs span
the lower frame while preserving contact, subject/look, and coworking context.
Turn `96` was useful secondary evidence for hands-on-hips/contact-centered
wording, but it did not improve the wide straddling silhouette as much.

Matrix update: sampler seed `2828282828` repeated the wide-thigh bridge
hierarchy across two visible women. Turns `208` and `212` were already valid
generic frontal cowgirl baselines with coherent coworking depth and centered
contact. Turns `209` and `213` improved the atlas silhouette by making the
thighs span left edge to right edge while preserving the viewer-underneath
body cues, upright torso, centered contact, and coworking setting. Turn `209`
was strongest because the viewer hands gripped both outer thighs; turn `213`
repeated the wide bridge with cleaner face/torso readability but weaker hand
visibility. Mirror the wide-thigh bridge hierarchy into the normal cowgirl
generator route provisionally. This is still a candidate route because the
baseline is already valid and another seed/source expansion should repeat the
refinement before promotion.

Generated-route validation: turn `216` used the patched normal cowgirl route
and preserved the full hierarchy: viewer lower abdomen and pelvis at the bottom
edge, hands gripping the outer thighs, wide thigh bridge across the foreground,
upright torso, centered contact below her belly, and coworking depth. Keep this
route patch provisional because it refines an already-valid baseline rather
than fixing a structural route failure.

Fresh-seed promotion: sampler seed `9191919191` repeated the patched normal
cowgirl route across turns `242`-`249` and two visible women. Turns `242` and
`246` used the generated-route wording directly and stayed valid: viewer-low
body ownership, lower abdomen/pelvis at the bottom edge, centered contact below
her belly, upright torso, coworking depth, and visible hand/thigh support.
Turns `242`, `243`, `244`, and `248` were the clearest atlas-like repeats for
the wide horizontal thigh sweep. Turn `247` was still valid cowgirl, but one
side did not span edge-to-edge as strongly. Treat the normal frontal cowgirl
wide-thigh bridge route as proven. Keep cowgirl-alt, reverse-cowgirl, and
reverse-cowgirl-alt separate because their atlas orientation is different.

### Cowgirl Alt / Flat Supine Low Angle

Cowgirl-alt should not be scored only by squat depth or contact clarity. The
atlas refs repeatedly show the man/viewer flat on his back while the woman is
mounted over him, and the room background proves the camera is low: ceiling,
upper walls, overhead lights, or high window/partition lines sit behind her
upper body.

Same-seed turns `97`-`100` showed the weak case. Low-squat wording produced
coherent woman-on-top images, and turn `99` improved supported-under-thigh
shape, but the background still read too much like a platform or high camera.
Do not accept that as cowgirl-alt just because contact and bent knees are
readable.

The accepted turn `104` hierarchy is:

- `viewer lies flat on his back underneath her`
- `lens sits low at the viewer's abdomen looking upward from his pelvis`
- `viewer abdomen and chest lie flat at the bottom edge`
- `woman mounted over him facing him`
- `knees bent wide and close to the camera`
- `ceiling lights, upper walls, or high glass partition lines behind her`
- `centered contact below her belly`

Patch only the cowgirl-alt route with this flat-supine/low-angle wording.
Normal frontal cowgirl can remain a higher, cleaner woman-on-top straddle when
the background still reads like a platform or elevated support.

### Reverse Cowgirl / Close Back-Hip Dominant

Generic reverse-cowgirl wording is not enough for Krea2. On seed `8989898989`,
turn `105` included `facing away`, viewer-underneath contact, and coworking
depth, but the woman still rendered as frontal cowgirl. The route needs the
back-facing straddle hierarchy before generic contact language.

For normal reverse cowgirl refs like `cowgirl_reverse/101_cowgirl_reverse.png`,
`cowgirl_reverse/106_cowgirl_reverse.png`, and
`cowgirl_reverse/1_cowgirl_reverse.png`, prefer:

- `woman sits on the viewer's pelvis facing away`
- `back, hips, and ass are the nearest largest shapes to the camera`
- `viewer thighs frame the lower corners`
- `centered contact sits directly between her thighs below her ass`
- `her thighs spread to either side of the viewer's hips`
- `partial over-shoulder glance` only after the back-facing straddle is locked

Turn `106` was the strongest same-seed candidate because it turned the frontal
baseline into a close back/hip-dominant straddle while preserving subject/look,
office depth, viewer-underneath legs, and centered contact. Turn `107` was a
valid secondary viewer-leg-frame variant, but a little farther back and more
upright. Turn `108` is useful secondary evidence for refs with an over-shoulder
glance.

Patch only the normal reverse-cowgirl route with this close back/hip hierarchy.
Keep reverse-cowgirl-alt separate: the alt atlas family is more upright and
seated, with the vertical back/shoulders and viewer hands or thighs forming a
cleaner seated straddle frame.

### Reverse Cowgirl Alt / Upright Seated Back-Facing

Reverse-cowgirl-alt is the upright seated branch, not just another close crop
of normal reverse cowgirl. The atlas refs such as
`cowgirl_reversere_alt/100_cowgirl_reversere_alt.png`,
`cowgirl_reversere_alt/101_cowgirl_reversere_alt.png`, and
`cowgirl_reversere_alt/102_cowgirl_reversere_alt.png` keep the woman's back and
shoulders vertical and readable above the hips, with the viewer's thighs and
often hands forming the lower seated straddle frame.

On seed `8989898989`, turn `109` showed that a subject-first upright baseline
already hits the route. Turn `110` was the strongest atlas-alt wording because
it added viewer hands on both hips while preserving the vertical back, centered
ass-over-pelvis placement, viewer thigh V-frame, centered contact, and coworking
depth.

Prefer:

- `POV upright reverse cowgirl back-facing penetration position`
- `woman sits upright facing away in a back-facing straddle`
- `back stays vertical and readable above her hips`
- `viewer hands hold both sides of her hips`
- `viewer thighs frame the lower corners`
- `centered contact remains visible below her ass`
- `office desks, glass partitions, plants, or tall-window light behind her upright back`

Use over-shoulder glance wording only as a secondary refinement. Turn `112` was
a strong over-shoulder variant, but the core route should first preserve the
upright seated back-facing silhouette.

### POV Doggy / Rear-Entry

For doggy-style POV, visible viewer thighs, lower torso, or pelvis can be
correct. Real POV references often show them. The useful target is not removing
the viewer body, but making the body cues read as a standing or crouched
first-person viewpoint instead of a vague seated pose.

To push the reference closer to a standing or crouched man looking down, use a
top-down rear-entry structure:

- `top-down standing POV doggy position from behind`
- `camera looks down over the viewer's extended hands onto the woman's raised hips`
- `woman is on all fours with chest low, forearms folded, cheek turned sideways`
- `rear-entry penetration visible between raised hips`
- `face and mouth remain far ahead, clearly separated from the penis`

Do not use visible shoes or lower legs as the standing cue. Seed `65` showed
that adding shoes/lower legs made Krea2 drift into oral contact and lose the
rear-entry geometry.

Do not over-prompt `viewer torso and thighs outside frame`; seeds `65` and
`6602` showed Krea2 still draws lower-body POV cues, and real references support
that. Prefer framing them as plausible foreground body cues rather than trying
to suppress them.

## Stronger-Control / Low-Priority Cases

Some atlas routes are useful to catalog but are not good prompt-only tuning
targets yet. Keep these out of the normal fixed-seed prompt queue until easier
pose families are covered.

### Sixty-Nine / Close Reversed POV

`pov_sixty_nine_close_reversed_oral` is currently the hardest and low-priority
atlas route. It should be treated as a pose/control image or image-guided route
first, not a normal prompt-only fixed-seed candidate.

The repeated atlas geometry is close and specific: the visible partner is
reversed over the viewer, hips closest to the camera, head and torso receding
away into the upper frame, and the viewer face or mouth anchoring the lower
foreground. Text-only prompting can collapse this into generic oral contact or
lose the reversed-over-viewer body arrangement.

## Style

Style should describe rendering, not camera mechanics.

Use style presets to choose between:

- natural photo,
- creator/social-media photo,
- documentary/direct-flash photo,
- cinematic realism,
- illustration/comic.

If a controlled camera is active, avoid style suffixes that imply a conflicting
camera such as `phone photo` or `handheld selfie`.

## Guide Update Format

When adding a new rule, include:

- observed prompt,
- observed image failure,
- edited prompt wording,
- image improvement or regression,
- generator path if known,
- final rule.