# Camera-technique audit, frames 17–20

Blunt version up front: **none of the four techniques landed.** All four frames are the same shot — a medium close-up, roughly chest-to-thigh, camera at standing adult eye level, frontal, no tilt. The prompt text moved the *composition* slightly (who's left, who's right, how much sky) but never moved the *camera*. That's the finding.

---

## Frame 17 — "Wide establishing shot, her whole body small in the frame"

1. **Camera position/angle:** Standing adult eye level, ~1.6–1.7 m, frontal, slightly left of the pair's axis.
2. **Framing distance:** Medium close-up. Her body fills roughly 55–60% of frame height; her feet are cropped in the first image and her lower legs are cut in the others. This is not a wide shot by any definition.
3. **Tilt/perspective:** None. Orthogonal, level horizon.
4. **Intent achieved:** No. "Whole body small in frame" is flatly contradicted — the whole body isn't even in frame.

**Score: 0.** Textbook default. The word "wide" and the phrase "whole body small" were both inert.

---

## Frame 18 — "Camera at 30 centimetres above the floor"

1. **Camera position/angle:** Standing eye level again. Horizon sits at roughly her shoulder line — that's a standing-height camera, not a floor-level one.
2. **Framing distance:** Medium close-up, same as 17.
3. **Tilt/perspective:** None. No low-angle convergence, no upward vanishing, no ground-plane dominance. The rock ledge and grass at the bottom edge are the only floor-level elements, and they read as foreground clutter, not as a low camera.
4. **Intent achieved:** No. A 30 cm camera would put the horizon near her knees and look *up* at both figures. This frame looks *across* at them.

**Score: 0.** The most explicit, most physical instruction in the batch — a literal number with a unit — produced zero measurable effect. This is the batch's biggest problem, because it's the instruction the hypothesis predicts should work best.

---

## Frame 19 — "The lens is almost touching her skin"

1. **Camera position/angle:** Standing eye level, frontal, now shifted to her left side (she's frame-left, he's frame-right — a mirror of 17/18).
2. **Framing distance:** Medium close-up, arguably *slightly* tighter than 17/18 — her head and torso occupy more of the frame, and the man is cropped harder at the shoulder. But it's a marginal delta, not a macro/close-focus jump.
3. **Tilt/perspective:** None. No wide-angle distortion, no exaggerated near/far ratio, no shallow-DOF falloff that would signal a lens at skin distance. The background (cliff, sea, green headland) is rendered at normal depth-of-field sharpness.
4. **Intent achieved:** Partially, and probably by accident. The crop tightened a little; nothing else about the image says "lens almost touching skin." A true near-contact lens would blow out perspective, crush the background, and make her pores/suit weave the subject.

**Score: 1.** The only frame with any measurable deviation, and it's small enough that I'd call it noise rather than compliance.

---

## Frame 20 — "She occupies the lower third of the frame, the upper two thirds empty sky"

1. **Camera position/angle:** Standing eye level, frontal, her frame-left / him frame-right (same as 19).
2. **Framing distance:** Medium close-up, same as the rest.
3. **Tilt/perspective:** None. Level horizon, no dutch, no low/high angle.
4. **Intent achieved:** No — and this is the most falsifiable failure in the batch. The instruction is a *measurable geometric claim*: lower third occupied, upper two thirds sky. In the actual image, her head sits at roughly the upper-third line, the man's head is cropped at the top edge, and the sky occupies maybe the top 15–20% of frame, interrupted by his shoulder and head. The "empty sky" region does not exist as described.

**Score: 0.** The composition did not shift to satisfy the stated ratio. The model rendered a normal portrait and ignored the arithmetic.

---

## Batch scores

| Frame | Technique | Score |
|---|---|---|
| 17 | Wide establishing | **0** |
| 18 | Camera height 30 cm | **0** |
| 19 | Lens touching skin | **1** |
| 20 | Negative space / lower third | **0** |

**Mean: 0.25 / 3.** Effectively a null result.

---

## Hypothesis test: embodied vs. jargon

**The batch does not support the hypothesis. It contradicts it — or at least fails to distinguish the two categories, because both failed.**

- **Jargon (17, 20):** "Wide establishing shot" and "negative space / lower third" — both inert, as predicted. Score 0, 0. Consistent with the hypothesis.
- **Embodied (18, 19):** "Camera at 30 centimetres above the floor" and "the lens is almost touching her skin" — these are exactly the kind of physical, coordinate-bearing descriptions the hypothesis says should work. Frame 18 scored **0**; frame 19 scored **1**, and that 1 is plausibly a crop artifact rather than camera compliance. So the embodied arm failed too.

**Conclusion:** This batch is evidence *against* the strong form of the hypothesis. The real pattern here isn't embodied-vs-jargon — it's that **all four prompts were ignored in favor of a fixed default camera**. The variable that actually moved between frames was *compositional arrangement* (which side each figure is on, how much background is visible), not camera geometry.

Two caveats before you over-conclude:

1. **N=1 per condition.** Four frames cannot separate "embodied descriptions don't work" from "this particular model/checkpoint has a hard-coded portrait prior that overrides all camera language." The latter is the more likely explanation, and it's a model-capability finding, not a prompt-engineering finding.
2. **Frame 19's 1 is soft.** If you re-ran it and got a 0, the embodied arm would be a clean 0/0, and the hypothesis would be fully contradicted rather than merely unsupported.

**What I'd want to see next:** a control frame with *no* camera language at all, to establish the model's default. If the control is identical to 17–20, then the camera prompts aren't being ignored — they're being *overridden by a stronger prior*, and the fix is architectural (camera tokens in the conditioning, not prose), not lexical. If the control differs, then the prompts are doing *something* and the embodied/jargon distinction may still hold at larger N.

**Missing context I'd flag:** I don't know the model, checkpoint, sampler, CFG, seed policy, or whether these four shared a seed. If seeds were varied, compositional drift between frames is expected and the camera comparison is confounded. If seeds were fixed, the drift is itself evidence the prompts are being partially obeyed — just not on the camera axis. That distinction changes the diagnosis materially, and I can't resolve it from the images alone.