Tune Krea2 POV doggy prompts

docs/krea2-pov-pose-atlas.md

@@ -0,0 +1,67 @@
+# Krea2 POV Pose Atlas
+
+Local reference root:
+
+`/media/unraid/davinci/Qwen_edit_lora/POV/dataset_v2`
+
+Use this dataset as the pose-geometry reference for POV prompt tuning. The pose
+folders contain rendered POV examples; matching `_control` folders contain the
+solo/control image for the same pose family. Ignore `bg` and `*_bg` folders for
+pose analysis; they are background plates without people.
+
+## Inventory
+
+| Family | Pose images | Control images | First sample |
+| --- | ---: | ---: | --- |
+| cowgirl | 63 | 63 | `5.cowgirl/100_cowgirl.png` |
+| cowgirl alt | 62 | 62 | `5.cowgirl_alt/101_cowgirl_alt.png` |
+| reverse cowgirl | 58 | 58 | `cowgirl_reverse/101_cowgirl_reverse.png` |
+| reverse cowgirl alt | 50 | 50 | `cowgirl_reversere_alt/100_cowgirl_reversere_alt.png` |
+| doggy | 57 | 57 | `doggy/101_doggy.png` |
+| doggy alt | 45 | 45 | `doggy_alt/100_doggy_alt.png` |
+| missionary | 74 | 74 | `missionary/101_missionary.png` |
+| missionary folded | 12 | 12 | `missionary_folded/16_missionary_folded.png` |
+| sixty-nine | 29 | 29 | `69/105_sixtynine.png` |
+| ballsucking | 25 | 25 | `ballsucking/101_ballsucking.png` |
+| blowjob laying | 42 | 42 | `blowjob_laying/101_blowjob_laying.png` |
+| blowjob side | 17 | 17 | `blowjob_side/103_blowjob_side.png` |
+| blowjob sitting | 27 | 27 | `blowjob_sitting/100_blowjob_sitting.png` |
+| blowjob top view | 17 | 17 | `blowjob_top_view/102_blowjob_top_view.png` |
+| boobjob | 11 | 11 | `boobjob/100_boobjob.png` |
+| handjob | 24 | 24 | `handjob/18_handjob.png` |
+| footjob | 2 | 2 | `footjob/59_footjob.png` |
+| fingering | 10 | 10 | `fingering/103_fingering.png` |
+| spread | 55 | 55 | `spread/100_spread_.png` |
+| ready | 19 | 19 | `ready/105_ready_.png` |
+| wand | 7 | 7 | `wand/106_wand_.png` |
+
+## Tuning Method
+
+For each pose family:
+
+1. Sample 5-10 pose images and 2-3 control images.
+2. Write a compact geometry summary using only repeated visual facts.
+3. Test one prompt variant with a fixed seed.
+4. Test the same wording on a second seed or character.
+5. Patch generator defaults only when the wording improvement repeats or the
+   generated prompt is structurally wrong before rendering.
+6. Record the evidence in `docs/krea2-prompt-guide.md`.
+
+## Confirmed Notes
+
+### Doggy / Rear-Entry
+
+Dataset references show that visible POV thighs, lower torso, or pelvis can be
+correct. They should be treated as natural foreground cues, not automatic
+failures.
+
+Better Krea2 wording:
+
+- `top-down POV doggy position from behind`
+- `camera looks down over the viewer's hands onto the woman's raised hips`
+- `woman is on all fours with chest low, forearms folded, cheek turned sideways`
+- `back arched, hips raised high toward the camera`
+- `viewer hands hold her hips with natural lower-body POV cues in the foreground`
+
+Avoid using visible shoes or lower legs as the standing cue. In seed `65`, that
+wording pulled Krea2 toward oral contact and weakened rear-entry geometry.

docs/krea2-prompt-guide.md

@@ -36,6 +36,11 @@ 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 seed,
@@ -140,6 +145,44 @@ Avoid:
 - 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.
+
+### 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.
+
 ## Style
 
 Style should describe rendering, not camera mechanics.

docs/sxcp-eval-loop.md

@@ -67,6 +67,29 @@ The request is sent on stdin. The command also receives:
 - Seed control/reproducibility
 - Overall Krea2 image usefulness
 
+## POV Pose Atlas
+
+Use `/media/unraid/davinci/Qwen_edit_lora/POV/dataset_v2` as the local
+reference atlas for POV pose geometry. The top-level pose folders contain real
+POV examples, and matching `_control` folders contain solo/control versions.
+Ignore `bg` and `*_bg` folders for pose rules; they are background plates
+without people. Treat the pose image folders as the primary source for body
+geometry; captions are optional and are not present for every folder.
+
+Suggested workflow:
+
+1. Choose one pose family, for example `doggy`, `doggy_alt`, `cowgirl`, or
+   `missionary`.
+2. Sample 5-10 real pose images and their control images.
+3. Write the repeated geometry as a compact prompt rule.
+4. Run one fixed-seed Krea2 prompt using that rule.
+5. Repeat on a second seed or character before changing generator defaults.
+6. If the prompt itself is structurally contradictory before rendering, patch
+   immediately and add a regression test.
+
+For POV doggy, the atlas shows that visible viewer thighs, lower torso, or
+pelvis can be correct. Do not treat them as automatic failures.
+
 ## Seed Contract
 
 The seed is transport metadata, not prompt text. When the graph emits a seed, an