# Blowjob Top-View 1024 Reference Pool Cue Expansion

Date: 2026-07-01

Canonical atlas folder:

```text
/media/unraid/davinci/Qwen_edit_lora/POV/dataset_v2/blowjob_top_view
```

Supplemental raw cue-expansion folder:

```text
/media/unraid/davinci/Qwen_edit_lora/POV/dataset_v2/1.original/blowjob_top_view_1024
```

Folder state:

- canonical folder: 17 PNG images
- supplemental raw folder: 27 PNG images
- all images are `1024x1024`
- raw-pool contact sheet created for review at `/tmp/blowjob_top_view_1024_contact.jpg`

Reference-pool report command:

```bash
python tools/krea2_atlas_refine_manifest.py --print-reference-pool-report --variant-key pov_blowjob_top_down_vertical_shaft --reference-pool-folder 1.original/blowjob_top_view_1024
```

Cue-review sheet command:

```bash
python tools/krea2_atlas_refine_manifest.py --print-reference-cue-review-sheet --variant-key pov_blowjob_top_down_vertical_shaft --reference-pool-folder 1.original/blowjob_top_view_1024
```

The cue-review sheet currently creates 27 blank review items: 17 canonical
curated references, all matched to raw counterparts by image id, plus 10
raw-only supplemental extras. Canonical rows include a
`reference_images_template`; raw-only rows leave that template empty until a
human review decides whether the extra frame is only cue-mining evidence or
deserves promotion into the curated catalog.

Filled cue-review sheet to sidecar-candidate draft command:

```bash
python tools/krea2_atlas_refine_manifest.py --print-reference-cue-candidate-draft --reference-cue-review-sheet-json /tmp/sxcp-reference-cue-review-filled.json
```

The candidate draft emits copyable `prompt_variant` objects only for reviewed
canonical rows with positive cues and a filled prompt-variant id. It skips
raw-only extras, noisy option/meta/negative cue wording, blank rows, and
duplicate ids so the raw pool cannot silently become promoted generator truth.

Candidate draft to same-stem sidecar authoring draft:

```bash
python tools/krea2_atlas_refine_manifest.py --folder /media/unraid/comfyui/output/CodexMCP-Atlas-Refine --subject-id atlas_refine_same_woman_001 --print-reference-cue-sidecar-author-draft --reference-cue-candidate-draft-json /tmp/sxcp-reference-cue-candidate-draft.json --variant-key pov_blowjob_top_down_vertical_shaft
```

Validate and apply only after confirming the target baseline deck is the one to
test:

```bash
python tools/krea2_atlas_refine_manifest.py --validate-reference-cue-sidecar-author-draft --reference-cue-sidecar-author-draft-json /tmp/sxcp-reference-cue-sidecar-author-draft.json
python tools/krea2_atlas_refine_manifest.py --apply-reference-cue-sidecar-author-draft --reference-cue-sidecar-author-draft-json /tmp/sxcp-reference-cue-sidecar-author-draft.json --folder /media/unraid/comfyui/output/CodexMCP-Atlas-Refine
```

Apply writes unscored sidecar prompt variants and checks the source prompt hash.
The next required action is fixed-seed MCP rendering and scoring, not catalog
promotion.

## Why This Pool Matters

The canonical folder is the preferred source for curated atlas references. The
supplemental raw folder shows the same pose family with more images, so it can
drive cue expansion when repeated visual axes are visible across references. It
should not be treated as proof that our current generated prompt can preserve
the same subject, coworking workspace, clothing ownership, or anatomy behavior.

Use this pool upstream of sidecar authoring:

1. Pick the nearest atlas cluster for the target variant, preferring the
   canonical `blowjob_top_view/` path when the frame exists there.
2. Extract positive cue axes from repeated visual evidence.
3. Write small `append_cues` against the current generated baseline and store
   the nearest atlas targets in `reference_images`.
4. Test with fixed sampler seeds through the MCP batch path.
5. Score generated results against the nearest atlas cluster and the current
   same-subject/workspace baseline before promotion.
6. Keep raw-only extras in the cue-review sheet until repeated generated
   evidence proves the cue belongs in a sidecar or catalog entry.
7. Convert only reviewed canonical rows to candidate sidecar snippets; then
   render and score them before seed selection.
8. Use the sidecar authoring draft to attach reviewed candidates to a baseline
   deck with prompt-hash drift protection.

## Observed Cue Axes

- `camera_pitch`: straight-down vertical, high oblique top-down, and tilted
  top-down variants all appear in the pool.
- `support_plane`: white lounge/chair surfaces, pale floor, carpet/rug, bed or
  blanket surfaces, wood floor, tile floor, and outdoor ground appear as visible
  plane anchors.
- `viewer_foreground_amount`: some refs use only a lower torso edge, while
  others show more thighs, feet, waistband, or abdomen mass.
- `partner_upper_body_stack`: the strongest top-view refs place face, eyes or
  eyelids, hair crown, shoulders, upper chest or neckline, and hand contact as
  the main partner stack.
- `hand_placement`: one hand at the base, both hands supporting the base, hands
  on floor/support, and hands on the viewer/body edge appear as alternate frame
  cues.
- `eye_direction`: direct eye contact, eyelids lowered, and open-mouth/gaze-up
  variants appear as expression axes.
- `clothing_anchor`: fitted tops, straps, shirts, or visible necklines can help
  anchor upper-body geometry when the crop supports them.
- `floor_furniture_evidence`: cropped support edges, floor seams, rugs, chair or
  furniture edges, and outdoor objects can carry scene identity without forcing
  deep room perspective.

## Current Top-View Rule From This Pool

For coworking prompt tests, translate workspace context into floor/support-plane
evidence and keep the shaft/contact line as the first pose anchor. The
user-highlighted atlas-22-style direction is:

```text
Straight-down male POV oral close-up. The centered shaft and mouth contact form
the main vertical axis from the lower foreground to the woman's face. The
woman's face, eyelids, hair crown, shoulders, upper chest, neckline, and one
hand stack around the shaft-contact axis. Viewer thighs and feet frame the
lower side edges. Tucked knees remain small side shapes on the floor. The
background reads as a flat pale floor and one cropped white lounge chair
surface, with shallow top-down room depth.
```

Manual same-seed calibration on 2026-07-01 confirms the anchor order. The
sidecars for `/media/unraid/comfyui/output/sxcp_accumulator/bwave_2/img_00135_.png`
through `/media/unraid/comfyui/output/sxcp_accumulator/bwave_2/img_00139_.png`
performed best when the prompt started with straight-down POV, then the shaft,
then the partner stack directly below the shaft, then mouth contact, and ended
with a sparse flat-floor background. Treat abdomen and room-depth cues as
secondary evidence; they should follow the shaft-contact axis instead of
leading the composition.

Because this finding depends on word order and removes the deep coworking-room
tail, test it as exact replacement text instead of appended cues. The validated
dry-run batch is:

```text
/tmp/sxcp_top_view_oral_shaft_anchor_exact_batch.json
```

It keeps the original baseline probe plus two exact-text candidates:

- `atlas22_shaft_contact_upper_stack_floor_plane`
- `atlas27_shaft_axis_between_feet_floor_anchors`

For clothed variants, keep clothing subject-owned:

```text
The woman wears a fitted white ribbed tank top; the tank-top neckline and
shoulders remain visible from above.
```

## Promotion Guard

Do not copy all 27 supplemental references into the live catalog variant. Keep
the catalog reference list curated, then use the raw pool to choose cue axes and
nearest visual targets. A generated candidate still needs fixed-seed image
evidence and visual scores for pose ownership, workspace continuity, clothing
visibility, subject identity, expression/eye control, anatomy/proportion, and
prompt noise before it can become seedable. Sidecar cue variants should carry
nearest visual targets like:

```json
{
  "id": "shaft_contact_upper_stack_floor_support",
  "text": "A same-subject straight-down male POV oral close-up. The centered shaft and mouth contact form the main vertical axis from the lower foreground to the woman's face. The woman's face, eyelids, hair crown, shoulders, upper chest, neckline, and one hand stack around the shaft-contact axis. The background reads as a flat pale floor plane with shallow overhead room depth.",
  "reference_images": [
    "blowjob_top_view/22_blowjob_top_view.png"
  ]
}
```