# Camera Technique Audit — Frames 01–08

**Scope note:** I'm working only from the eight images and the manifest you supplied. No repository files, no generation logs, no seed/CFG metadata — so I cannot attribute any difference to seed drift vs. prompt, and I cannot verify that the pose is *literally* identical across frames (it visibly isn't — see the pose-drift caveat at the end).

---

## Per-frame findings

**01 — T01 "Low angle shot"**
1. **Camera:** Level, slightly below her eyeline at most. Lens sits roughly at her chest/shoulder height, frontal-left.
2. **Framing:** Medium — head to upper thigh, both figures' torsos in frame.
3. **Tilt:** None. Straight-on, horizon level.
4. **Intent:** Not achieved. This is the baseline eye-level frontal shot. The "low angle" token did nothing.
**Score: 0**

**02 — T02 "Low angle shot, 24mm wide lens"**
1. **Camera:** Still essentially level, marginally lower than 01 — lens around her sternum. Frontal-left.
2. **Framing:** Medium-tight, head to knee. Slightly wider apparent FOV than 01, consistent with the 24mm token having *some* effect on lens character.
3. **Tilt:** None.
4. **Intent:** The lens token landed; the angle token did not. Camera position is unchanged from 01. This is a lens test, not an angle test.
**Score: 1** (deviation is optical, not positional)

**03 — T03 "The camera is on the floor looking up at her"**
1. **Camera:** Genuinely low — lens is at or below the level of her hip/seat, looking up. There is a visible camera prop sitting on the rock at bottom-left, which the model appears to have *rendered as an object* rather than obeyed as a viewpoint instruction.
2. **Framing:** Medium-wide — head to feet, full seated figure plus the man's torso.
3. **Tilt:** Slight upward convergence on the man's torso; her body reads foreshortened from below.
4. **Intent:** Partially achieved. The upward angle is real but mild, and the model's literal-minded rendering of "camera on the floor" as a physical camera is a tell that it treated the sentence as *scene content*, not *viewpoint*.
**Score: 2**

**04 — T04 "Shot from the perspective of someone lying on the ground"**
1. **Camera:** This is the strongest deviation in the set. Lens is low and frontal, looking up at her from roughly ground level — and critically, the man's body is *occluding* the bottom of frame as a foreground mass, which sells the "someone lying there" POV.
2. **Framing:** Tight on her torso and hips, head cropped at the top by his chin/chest. His arms bracket the frame.
3. **Tilt:** Strong upward angle; her body looms. Foreground occlusion creates real depth.
4. **Intent:** Achieved. The POV framing (his body as the near-field) is exactly what "perspective of someone lying on the ground" should produce.
**Score: 3**

**05 — T05 "Her feet are closest to the lens and her head is far away"**
1. **Camera:** Low, frontal, positioned near her feet — but the depth cue is weak. Her raised leg is in the near field and her head is smaller/further, so the *relative* scale instruction was partially honored.
2. **Framing:** Medium — head to foot, full body diagonal.
3. **Tilt:** Mild foreshortening on the near leg; not a dramatic wide-angle stretch.
4. **Intent:** Partially achieved. The compositional relationship is present but understated; this reads closer to a normal three-quarter shot than a true feet-first perspective.
**Score: 1**

**06 — T06 "Extreme close-up of her face from below"**
1. **Camera:** Low and close, angled up at her face. Her chin/neck dominate, nostrils and jaw underside visible — a real below-face angle.
2. **Framing:** NOT extreme close-up. It's a medium shot: head, shoulders, torso, and the man's chest all in frame. The "extreme close-up" half of the instruction was ignored.
3. **Tilt:** Upward angle on the face is genuine.
4. **Intent:** Half-landed. The "from below" worked; the "extreme close-up" did not. The model prioritized the angle over the crop.
**Score: 2**

**07 — T07 "Dutch angle, the horizon tilted sharply"**
1. **Camera:** Level, frontal-left — same family as 01/02.
2. **Framing:** Medium, head to thigh.
3. **Tilt:** **None.** The sea horizon behind her is dead level. No roll whatsoever.
4. **Intent:** Failed outright. "Dutch angle" and even the explicit "horizon tilted sharply" produced zero camera roll. This is the clearest null result in the set.
**Score: 0**

**08 — T08 "Over-the-shoulder shot from behind her left shoulder"**
1. **Camera:** Frontal-left, roughly at her shoulder height. It is *not* behind her shoulder — the lens is in front of her, seeing her face three-quarter.
2. **Framing:** Medium — head to hip, her near shoulder and the man's chest.
3. **Tilt:** None.
4. **Intent:** Failed. The "over-the-shoulder from behind" framing requires the camera to be behind her head with her shoulder as foreground; instead we get a standard frontal three-quarter. The model read "over-the-shoulder" as a *compositional label*, not a *camera placement*.
**Score: 0**

---

## Verdicts

**Techniques that worked (score ≥2):**
- **T04** — "perspective of someone lying on the ground" (3). The only unambiguous success.
- **T03** — "camera is on the floor looking up" (2). Real but mild; partially literalized as a prop.
- **T06** — "extreme close-up of her face from below" (2). Angle landed, crop did not.

**Techniques that defaulted to eye-level frontal (score ≤1):**
- **T01** — bare "low angle shot" (0). Total default.
- **T02** — "low angle shot, 24mm" (1). Lens changed, angle didn't.
- **T05** — "feet closest, head far" (1). Understated; near-default.
- **T07** — "Dutch angle, horizon tilted sharply" (0). Total default; no roll at all.
- **T08** — "over-the-shoulder from behind her left shoulder" (0). Total default; camera stayed in front.

---

## Pattern

**Yes, and it's stark.** The correlation is not "explicit vs. bare token" — it's **whether the sentence describes a physical body in a physical location** versus **whether it names a cinematographic term.**

- **T04** ("someone lying on the ground") and **T03** ("camera is on the floor") are *embodied* descriptions: they place a human body or object at a spatial coordinate, and the model can simulate that viewpoint. Both scored ≥2.
- **T01** ("low angle shot"), **T07** ("Dutch angle"), **T08** ("over-the-shoulder shot") are *jargon labels*. The model treats them as style tags with weak or zero geometric force. All scored ≤1, and T07/T08 scored 0 despite being maximally explicit ("horizon tilted sharply," "from behind her left shoulder").
- **T06** is the instructive middle case: "from below" (embodied) worked; "extreme close-up" (jargon) did not. Within a single sentence, the embodied clause landed and the jargon clause was dropped.
- **T02** and **T05** confirm the same split: "24mm wide lens" is a *lens* parameter and produced a lens effect; "low angle" and "feet closest" are weak spatial claims and produced weak spatial effects.

**Blunt conclusion:** explicit wording alone does not help. What helps is **describing the camera as a physical object or a person's eye at a specific location** ("on the floor," "someone lying on the ground"). Naming the shot type ("low angle," "Dutch," "over-the-shoulder") is close to inert in this pipeline. If you want a Dutch tilt, do not say "Dutch angle" — say "the camera is rolled thirty degrees to the side, the sea horizon cuts diagonally across the frame." If you want over-the-shoulder, say "the camera is behind her head, her left shoulder fills the near foreground."

---

## Caveat on the premise

The brief says the pose and act are identical across all eight. They are **not** — the girl's leg configuration, seating surface, and the man's grip change materially between frames (01 has her against a rock wall with legs down; 03 has her seated on a stone ledge with a camera prop; 04 has her upright with legs apart and his hands on her hips; 05 has a raised leg; 08 has her leg hooked over his forearm). So some of the apparent "camera variation" may be **pose variation leaking through**, which would contaminate the attribution. Frames 01, 02, 07, and 08 are the most pose-consistent cluster and all four are frontal/level — that cluster is the cleanest evidence for the "jargon is inert" finding. Frames 03, 04, 05, 06 differ in pose *and* camera, so their scores are less cleanly attributable to technique alone.