# Camera-distance test — method and result

Run: 2026-10-25. 2 comic pages × 3 runs = 6 pages, plus 2 pilot pages.

## Why this was run

/consistent-character-ai said the reference sheet is applied the same way at
every camera distance, so a distant figure has far less face to match and
"that is where identity is most likely to slip". That was a reasoned claim
with no evidence behind it. This test checks it.

## What was generated

Two four-panel comic pages that read as one short opening: Mira returns to
the canal town she left six years ago.

- **Page A, "Arriving at dusk"** steps the camera closer panel by panel:
  extreme wide (her figure about 8% of the panel), wide (about 30%), medium
  (waist up), close-up. Three text elements.
- **Page B, "Night market"** is the hard case: a crowd shot (about 12%,
  behind passers-by), a back three-quarter wide shot, a medium two-person
  panel with a stall owner, a close-up. Three text elements.

The exact prompts are published alongside this file: `prompt-A.txt` and
`prompt-B.txt`. All three runs of a page used the identical prompt.

## How it mirrors the product

- One image per page, from the page model (GPT Image 2 edit), 1024×1536.
- The **whole** reference sheet is attached: `persona-test/persona-A-from-photo.webp`,
  the same sheet behind the 8/8 photo run.
- Alongside it, a written description of the **face only**. The outfit is
  not written down; the sheet's full-body half carries it, as in the product.
- Style matched to the sheet's own style, so a style mismatch could not
  pass for identity drift.

**The one difference from the product:** there is no story-planning step.
The page prompts were written by hand.

## Rules held throughout

- Camera distance pinned by her height in each panel.
- Every published run is a **first attempt**. Nothing was rerun or picked.
- Panels too small to show a detail are marked "not visible at this size",
  never counted as a pass.

## Pilots (prompt development, not counted)

The first prompt produced one page of each design (`pilot-A`, `pilot-B`).
The model **enlarged her** in the wide panels — the crowd shot came out at
about 45% of panel height instead of the 25% asked for, front and centre,
which made the hard case easy. The prompt was tightened (explicit "she
must stay small", far third of the street, larger people in front of her)
and the wording for the beauty mark was changed from "her left eye" to
"exactly where it is in the sheet", because "left" was ambiguous. The
pilots are published but not scored.

## Scoring

Each panel showing Mira, against the reference sheet: face, hair (long,
dark, wavy, centre part), beauty mark, pearl studs, outfit (ochre jacket,
matching trousers, cream t-shirt, white sneakers).

| Distance | Panels | Face | Beauty mark | Pearl studs | Outfit |
|---|---|---|---|---|---|
| Close-up | 6 | 6/6 same person | 6/6 | 6/6 | 6/6 |
| Medium | 6 | 6/6 | 3/3 where visible; 3 turned away | 5/6; 1 hidden by hair | 6/6 |
| Wide | 6 | 3/3 plausible where shown; 3 are back views | not visible at this size | not visible | 6/6 |
| Extreme wide / crowd | 6 | not resolvable at this size | not visible | not visible | 6/6 |

Lettering: 18 of 18 text elements spelled correctly, with tails on the
right speaker.

## Findings

1. **At a distance the face did not slip — it stopped being checkable.**
   In every extreme-wide and crowd panel the face is a handful of pixels.
   She is still recognisable, from her outfit and hair: the full-body half
   of the sheet is what carries identity at that size.
2. **Look-alike extras in crowds.** In B run 2's crowd panel a woman in the
   foreground wears an ochre jacket with long dark wavy hair — effectively a
   second Mira. B runs 1 and 3 have milder versions (tan jackets, long dark
   hair). The reference bled into the background cast. Count: 1 clear case,
   2 partial, of 3 crowd panels.
3. **Hair texture drifts before the face does.** The sheet's hair is wavy
   and dark brown. In the market close-up and medium panels it came out
   straighter and blacker in 2 of 3 runs (B2, B3), under red lantern light.
   The face held in all of them.
4. Close-ups and medium shots held completely.

## What this is not

Six pages is a small sample. It shows what happened on these six first
attempts, not a measured rate. It is one character, one style, one model.