# Camera Technique Audit — Frames 01–08

**Blunt summary up front:** of eight techniques, **one** landed strongly (T08), **one** partially (T04), and the other six collapsed into variations of the same eye-level frontal portrait. The model is not obeying camera-position language; it is obeying *framing* language and *narrative* language, and ignoring *angle* language.

---

## Per-frame

**01 — t01_bare_low_angle ("Low angle shot")**
- Camera: essentially eye-level, frontal, maybe a hair below chin height. Lens is at her sternum/chin line, not below.
- Framing: mid-thigh to top of head; standard 3/4 portrait crop.
- Tilt: none.
- Intent: **failed.** The horizon is level, the verticals are level, and there is no upward convergence on her body. A true low angle would foreshorten her legs and put the horizon low in frame or below the frame edge.
- **Score: 0**

**02 — t02_low_angle_24mm ("Low angle shot, 24mm wide lens")**
- Camera: eye-level frontal, dead center. If anything slightly *above* eye level — we're looking marginally down at her.
- Framing: full torso, crotch-to-crown, tight vertical crop.
- Tilt: none. No wide-angle distortion visible; her proportions are normal, so the "24mm" did not manifest as perspective stretch.
- Intent: **failed.** Adding a focal length to a failing angle token did not rescue it. The lens spec was ignored as thoroughly as the angle.
- **Score: 0**

**03 — t03_camera_on_floor ("The camera is on the floor looking up at her")**
- Camera: this is the interesting one. The lens is *low* — roughly knee/shin height, and there is a visible physical camera body sitting on the grass in the bottom-left foreground. But the *view* is not a floor-level upward view: we see the top of her feet and her full standing body without dramatic foreshortening. It reads as a camera placed on the ground but pointed at her chest, not tilted up.
- Framing: full body, feet to head, with foreground ground plane.
- Tilt: none; horizon level.
- Intent: **partially achieved** — the *object* "camera on the floor" was rendered literally, but the *perspective* it implies was not. This is the classic failure mode: the model drew the noun instead of adopting the viewpoint.
- **Score: 1**

**04 — t04_photographer_lying ("Shot from the perspective of someone lying on the ground")**
- Camera: low, near ground, and crucially there is a **visible pair of bare feet and blue-clad legs in the foreground bottom** — a body lying down, rendered as an actual foreground element. The lens is at roughly ankle height.
- Framing: knees to head; her feet are cropped out, and the foreground body occupies the bottom edge.
- Tilt: none.
- Intent: **partially achieved, and better than 03.** The narrative "someone lying on the ground" produced a genuine ground-level viewpoint *and* a foreground body. But it's still not a dramatic upward angle — her head is not looming, the horizon is still level, and the body reads as a prop rather than a POV.
- **Score: 1** (arguably 1.5 — the strongest of the low-angle attempts)

**05 — t05_feet_closest ("Her feet are closest to the lens and her head is far away")**
- Camera: **this is the closest thing to a real low angle in the set.** The lens is low, near her feet, and there is genuine perspective recession — her feet are large in frame, her legs converge upward, her head is smaller. The cliff edge behind her drops away, reinforcing the height.
- Framing: full body, feet at bottom edge, head near top.
- Tilt: none, but the *perspective convergence* is real.
- Intent: **achieved, mostly.** This is the one frame where the model actually committed to a viewpoint rather than a label. It's not extreme, but the foreshortening is unambiguous. Note: this succeeded because the prompt described *spatial relationships* ("feet closest, head far") rather than naming an angle.
- **Score: 2**

**06 — t06_extreme_close_below ("Extreme close-up of her face from below")**
- Camera: **eye-level or slightly above**, frontal, tight. The "from below" was completely ignored. We are looking at her face straight on, maybe a touch down.
- Framing: head and shoulders, tight crop — the "extreme close-up" part landed.
- Tilt: none.
- Intent: **failed on the angle, succeeded on the framing.** The model honored "extreme close-up" and discarded "from below." Her chin is not lifted, her nostrils are not visible, there's no upward convergence.
- **Score: 1** (framing credit only; angle is a 0)

**07 — t07_dutch_tilt ("Dutch angle, the horizon tilted sharply")**
- Camera: eye-level frontal, mid-thigh to head.
- Framing: standard portrait crop.
- Tilt: **none.** The horizon behind her is level. The cliff edge is level. There is no rotation of the frame whatsoever.
- Intent: **failed.** "Dutch angle" as a bare token did nothing. The model produced a perfectly upright, level-horizon portrait.
- **Score: 0**

**08 — t08_over_shoulder ("Over-the-shoulder shot from behind her left shoulder")**
- Camera: **this one landed.** The lens is behind and to her left, looking past her shoulder toward the sea. Her body is angled away, we see her back-left shoulder and the turn of her head toward camera. The composition is genuinely over-the-shoulder.
- Framing: waist-up, with her shoulder occupying the left foreground and the sea/cliff filling the right.
- Tilt: none, but the *positional* change is unambiguous.
- Intent: **strongly achieved.** This is the only frame where the camera position is unambiguously not eye-level frontal.
- **Score: 3**

---

## Verdicts

**Techniques that worked:**
- **T08 over-the-shoulder** — 3. Fully landed.
- **T05 feet-closest** — 2. Real perspective recession; the best low-angle result.
- **T04 photographer-lying** — 1. Ground-level viewpoint with foreground body, but not dramatic.
- **T03 camera-on-floor** — 1. Literal prop rendered, perspective not adopted.

**Techniques that defaulted to eye-level frontal:**
- **T01 bare low angle** — 0.
- **T02 low angle + 24mm** — 0.
- **T06 extreme close-up from below** — 1 (framing only; angle defaulted).
- **T07 Dutch tilt** — 0.

---

## Pattern

Yes, and it's stark:

1. **Bare angle tokens fail.** "Low angle shot" (T01), "Dutch angle" (T07) — both produced default eye-level frontal portraits. The model treats these as *style adjectives*, not *camera instructions*.

2. **Adding a lens spec does not help.** T02 added "24mm wide lens" to the same failing token and got the same failure. Focal length is decorative unless the viewpoint is already established.

3. **Spatial-relationship descriptions work.** T05 ("feet closest to the lens, head far away") succeeded because it described *what the image should look like* in terms of relative scale and distance, not what the camera should be called. This is the single most reliable technique in the set.

4. **Narrative/embodied descriptions partially work.** T03 and T04 ("camera on the floor," "someone lying on the ground") produced literal props and ground-level viewpoints — the model rendered the *noun* rather than adopting the *POV*. Useful, but not sufficient.

5. **Named compositional setups work best.** T08 ("over-the-shoulder from behind her left shoulder") is a *named shot type* with a *specified side*, and it landed cleanly. The specificity of "from behind her left shoulder" gave the model a concrete spatial anchor.

6. **"From below" is the weakest modifier.** Both T01, T02, and T06 used it in some form and all failed. The model appears to have a strong prior toward eye-level portraiture that overrides explicit "below" language unless paired with a spatial-relationship description (as in T05).

**Recommendation for the technique test:** replace bare angle tokens with spatial-relationship sentences ("her feet fill the bottom of the frame and her head is small at the top"), and for Dutch tilt, describe the *result* ("the horizon runs diagonally across the frame") rather than naming the technique. T08's success suggests that naming a shot type *plus* specifying the side is the strongest prompt pattern observed.