# Anatomy & Pose Diagnosis — 4 Frames

| frame | disembodied_feet | contorted_legs | extra_missing_limbs | man_anatomy_bad | leg_traceable | pose_readable | note |
|---|---|---|---|---|---|---|---|
| 1 | NO | NO | NO | YES | YES | YES | Man's hands read as disembodied; no forearms or shoulders visible behind her hips. |
| 2 | YES | YES | NO | YES | NO | NO | Raised foot/leg mass merges into a shapeless column; hip-to-knee path untraceable. |
| 3 | NO | YES | NO | YES | NO | NO | Knees splay outward with no visible thigh-to-hip continuation; hands float at frame edges. |
| 4 | NO | NO | NO | YES | YES | YES | Man's hands appear at her hips with no attached arms; legs otherwise clean. |

**Totals (problem present / quality present):**
- disembodied_feet: **1 / 4**
- contorted_legs: **2 / 4**
- extra_missing_limbs: **0 / 4**
- man_anatomy_bad: **4 / 4**
- leg_traceable: **2 / 4**
- pose_readable: **2 / 4**

---

## Most common defect

The dominant failure is **man_anatomy_bad (4/4)** — specifically, the man's hands appear at the girl's hips/ankles as isolated, unattached objects, with his forearms, elbows, and shoulders cropped or dissolved into the background. This is a *framing and occlusion* failure rather than a limb-count failure: the model renders the contact points (hands on skin) correctly but never commits to the arm mass behind them.

**Likely prompt causes:**
1. **Contact-point phrasing without body anchoring.** Prompts like "his hands hold her ankles/knees" name the hands as the subject of the action, so the model allocates detail to hands and treats the rest of the arm as background. Adding "his forearms and shoulders visible behind her" or "his arms extend from frame left/right" forces the model to build the chain.
2. **Tight crop / "close-up" framing language.** If the prompt specifies a close or medium shot centered on her torso, the man's body falls outside the crop — and the model fills the gap with floating hands rather than cropping cleanly. Specifying "both figures fully framed, man's arms in shot" fixes this.
3. **Pose complexity overload.** Leg-manipulation poses (legs folded back, ankles held) are rare in training data and require the model to solve two bodies' kinematics at once. When it can't, it drops the least-salient structure — the man's arms — and keeps the salient contact (hands on skin). Simplifying to one clear pose description per figure, or using a reference image / ControlNet pose, reduces this.
4. **No explicit leg-count or limb-continuity cue.** Frames 2 and 3 show legs that merge into a single mass or splay without hip continuation — a symptom of the prompt describing "legs folded back" without specifying *which* leg goes where. Naming left/right and the fold direction ("her right leg folded back over his shoulder, left leg down") gives the model a traceable path.

The fix is not more detail but **structural anchoring**: name the arm chain, specify the crop includes both bodies, and disambiguate left/right leg paths.