# MJ Contact Sheet — one brief, many angles, proven attribution A comparison sheet for the Midjourney pipeline: one brief rendered at N different camera angles, laid out as a single contact sheet with one tile per variant, each tile labelled with the clause that produced it. The tool is `lib/mj-contact-sheet.js`. The shipped example is `assets/mj-sheets/red-coat-underpass-12/` and the page that presents it is `pages/mj-sheets.html`. --- ## What it does ```bash node lib/mj-contact-sheet.js --pack red_coat_underpass_12 \ --out assets/mj-sheets/red-coat-underpass-12/contact-sheet.png \ --tiles assets/mj-sheets/red-coat-underpass-12/tiles \ --manifest assets/mj-sheets/red-coat-underpass-12/sheet-manifest.json \ --title "Red Coat Underpass" ``` | Flag | Meaning | |---|---| | `--pack ` | directory under `mj-output/` (required) | | `--out ` | the sheet PNG (default `mj-output//contact-sheet.png`) | | `--tiles ` | also save each tile as its own file (`prompt-NN.webp`) | | `--manifest ` | write the provenance manifest | | `--title ` | heading text (default: derived from the pack slug) | | `--cols` / `--rows` | grid shape (default 4 × 3 = 12 up) | | `--thumb ` | tile width in CSS px (default 480) | | `--slot ` | which grid cell of the 4-up to use by default (default 0) | | `--slots "2=1,5=3"` | per-variant cell override, keyed by prompt number | Exit 0 sheet written · 2 bad input · 3 render failure. --- ## Why a browser and not `curl` The Midjourney CDN (`cdn.midjourney.com`) answers curl and headless clients with a Cloudflare bot challenge — HTTP 403. A real browser gets the pixels. The builder therefore drives Chrome **headed** (the same constraint the web backend honours), renders the sheet as HTML, and screenshots it. The same reason the tiles are exported through the page (`fetch` + `FileReader` in-browser) rather than by a separate download step: the browser is the only client the CDN serves. --- ## The four things that can silently corrupt a sheet ### 1. Wrong-job attribution The imagine feed is a **shared timeline**. The account owner's own concurrent jobs sit in it, with their own prompts and their own images. A matcher that reads a rendered card's text and looks for the prompt inside it will happily match the owner's job. That is not hypothetical: it happened on 2026-10-21, and 6 of 12 tiles in the first sheet were the owner's portrait rather than the brief. **The fix** is `lib/mj-feed.js`: the sheet's task ids are re-derived from the feed's own `full_command`, and only plain `16:9` `v8.2` jobs are accepted. The DOM matcher was deleted, not demoted — keeping an unsound matcher as a "fallback" is how it came back. `lib/mj-web.test.js` asserts it is not exported. **What this means for you:** a task id on a sheet is only as good as the gate that produced it. If you swap the gate, re-check the tiles by eye. ### 2. Flags that drifted A job enqueued as `--ar 4:3 --stylize 0` instead of `--ar 16:9 --s 100` is a *successful* job with the wrong parameters — no error, just a different picture. The drift came from the account's ambient UI state (`localStorage.settings_v30`), not from the prompt. The feed gate therefore checks the **enqueued command**, not just the prose: the requested `--ar` must be present with the requested value, and `--edit` / `--cref` / `--oref` must be absent. `--stylize` is deliberately **not** strict — Midjourney omits it when it equals its default, so demanding it would reject good jobs. ### 3. A caption that does not match its tile The caption is not a description of the result; it is the clause from the prompt that produced the tile. It has to be *derived*, not written by hand, or it drifts. `variantLabel()` reads the clause back out of the prompt — the text between the shared head (`…underpass at night,`) and the shared tail (`one overhead sodium streetlight…`) — so the caption and the image are provably in sync. **A bug worth knowing about:** the caption cap used to truncate at the *first* separator, so `overhead mirror reflection, red coat doubled in glass, muted 35mm documentary` collapsed to `overhead…`. It now cuts at the *last* separator in the kept span. `lib/mj-contact-sheet.test.js` pins this. --- ### 4. A sheet that no longer came from its pack The three failures above happen *while* a sheet is being built. The fourth happens afterwards: someone edits the pack, re-renders one variant, or regenerates the manifest from a different run — and the shipped artifact quietly stops being the output of the pipeline. **The fix is a guard on the artifact, not on the code.** A test on synthetic input proves the builder behaves; it does not prove the shipped sheet is what the builder produced. The gap between the two is where drift lives. `lib/mj-contact-sheet.test.js` (section 5) therefore walks every directory under `assets/mj-sheets/` and pins the whole chain of custody, **data-driven** — a second pack is covered with no edit to the test. It asserts: - the pack still exists and still parses with the submitter's own `extractPrompts`; - the pack yields exactly the manifest's variant count; - every prompt is `16:9`; - it is **one brief with twelve variants** — proved by longest-common-prefix and longest-common-suffix, not by hardcoding marker strings; - every caption is the clause that produced its tile, and every caption is a prefix of its own clause; - the twelve task ids are distinct, and every tile's URL embeds its own task id; - every tile's URL names the **same cell of the 4-up** the manifest records as its `slot` — the manifest and the URL are written by different parts of the builder, so they can drift apart, and a manifest claiming "slot 2" over a slot-1 tile misreports the comparison the sheet exists to make; - every tile ships as a **local real webp with real pixels**, and **no two tiles are the same image**; - the sheet PNG is a real PNG, large enough for its grid. **It reads tracked files only.** `mj-output/` is gitignored, so a guard that needed the metas would be skipped in CI — which is exactly where a broken sheet would be published. **The guard was proven by making it fail.** Eight mutations (mangled caption, mangled task id, duplicated task id, duplicated tile, mangled pack clause, mangled count, missing tile, stubbed PNG) were each caught by the guard that should catch them. A guard that never fails is decoration. --- ## Per-variant slot selection Midjourney returns a **4-up grid** per prompt, and the four cells are not four renditions of one framing — they are four different takes on the brief. A prompt that says "overhead crane shot straight down" often yields one genuinely top-down cell and three eye-level ones. `--slots "2=1,5=3"` picks a specific cell per variant (the key is the prompt's 1-based number). Unlisted variants use `--slot`. The chosen cell is recorded in the manifest as `slot`, so the sheet says which of the four it is showing. --- ## When Midjourney ignores the angle It happens. Two of the twelve prompts came back as eye-level and level respectively — the crane and the dutch tilt were both ignored, silently, in a job that reported success. Rewording helped both, and the rewrites are what shipped: - `high overhead crane angle looking straight down` → `overhead crane shot straight down, bird's eye view` - `dutch tilt three-quarter view` → `dutch tilt, camera rolled 30 degrees, tilted horizon` **Verify by looking.** The pack's own check is to render every cell of a suspect job and pick the closest one; if no cell is close, reword and re-render. A job that "succeeded" is not evidence that the angle landed. ### The third case: a caption that lies about the camera Variant 09 was the shipped sheet's worst lie, and it survived two audits before anyone looked. Its caption read *overhead mirror reflection, red coat doubled in glass*; the tile was a frontal shot with the reflection in a mirrored ceiling. The camera was never overhead. The pack's own check is the first step, and it is worth doing exactly as written — render **every** cell of the suspect job and pick the closest one: - the job behind tile 09 (`8fd1fb8d`) has four cells; - put all four next to a **known-overhead** tile and ask which one is shot from above; - the answer was *none of them*. Measured geometry is not a substitute for that look. A red-mass profile (the share of red pixels in each third of the frame) called one cell "top-weighted, like the known overhead" — it was the wrong cell. The red mass in a wet-floor scene is a reflection, and a reflection moves the number without moving the camera. CLIP image-to-image similarity is no better: ranked against the known overhead tile it put an eye-level cell first. Two instruments, two wrong answers, one lesson — **the sheet's audit is a look, not a measurement.** So: reword, re-render. `looking down from a high angle, red coat doubled in the wet floor` delivered high-angle cells (verified by looking, one cell at a time). ### The condition the tiles were rendered under The re-render came back **three times darker** than the other eleven tiles, and the reason was not the angle. The twelve original jobs carried the account's profile **plus four moodboards** (`--profile brvazka mhw6dpq xr39csb det21e6 67lkfw3`); a job submitted the next day carried `--profile brvazka` only. The moodboards are still on the account — the composer's Personalize panel just no longer expands the profile into them, so a new job inherits the profile and nothing else. A controlled A/B settles it: the same reworded prompt, the same flags, the profile codes typed into the prompt text instead of injected by the composer. | job | profile in the command | mean luminance | red pixels | |---|---|---|---| | original set (11 tiles) | profile + 4 moodboards | 0.23 – 0.39 | 28k – 265k | | re-render, composer only | profile | 0.08 – 0.13 | 6k – 14k | | re-render, codes in the prompt | profile + 4 moodboards | 0.26 – 0.36 | 28k – 265k | Two things follow. **A typed `--profile` wins cleanly.** The composer still injects its own `--profile brvazka` when the prompt carries one, and the command that reached Midjourney had **exactly one** `--profile` — the typed one, with all five codes. No duplicate flag, no need to touch the account's toggle. (The account's style codes are deliberately **not** written into the pack: they are the account's own identifiers, and the pack is tracked. Read them off any older job's `full_command` instead — the feed keeps them.) **A comparison sheet needs one rendering condition.** The shipped tiles were all rendered under the profile and its moodboards; the re-render was matched to them. A tile that is dark because it came from a different account setting is not a comparison, it is a different picture. --- ## The shipped example `assets/mj-sheets/red-coat-underpass-12/` ``` contact-sheet.png 1990 × 996 — the 12-up tiles/prompt-01..12.webp 1143 × 640 — the individual tiles sheet-manifest.json provenance: label, task id, slot, CDN url per variant ``` The brief: *a lone figure in a saturated red coat standing three-quarter turned in an empty concrete underpass at night, hip-height 35mm film still, one overhead sodium streetlight as the only key, wet floor catching thin reflections, deep unlit tunnel falling away behind, 35mm lens wide open, fine natural film grain, muted palette, no text* — `--ar 16:9`. Twelve bullets share the head and tail byte-for-byte and vary only the clause between them. That is what makes the sheet a comparison rather than twelve unrelated pictures. The pack lives at `assets/mural-prompts/red-coat-underpass-12.md` (12 prompts — an exception to the 5–10-per-file convention, noted in the pack header). Eleven tiles are cell 0 of their job. **Variant 09 is cell 3 of `1f583a22`** — the re-render described above; the manifest records the slot, and the tile's own URL names the same cell. --- ## Tests ```bash npm run test:mj-contact-sheet # 49 checks, no browser, no network ``` Covers the caption derivation (including the word-boundary truncation), the grid-cell picker, the per-variant slot map, the HTML builder, and the **shipped artifact's chain of custody** (section 4 above). The reconcile tool has its own suite: ```bash npm run test:mj-feed-reconcile # 43 checks, no browser, no network ``` It covers the half of `lib/mj-feed-reconcile.js` that **writes** the pack's metas — the repair path for the pipeline, and therefore the path that most needs a test. Among other things it pins the `taskIdChanged` contract: two ids are the same job when one is a prefix of the other. That tolerance is deliberate — the comparison it replaced truncated the old id to 8 characters and compared it against a full uuid, so the "changed" flag was **always** true, and every line of the output said `CHANGED` even when nothing had changed. `e2e/mj-sheets.spec.mjs` covers the shipped page: the sheet image, the twelve tiles, each caption tied back to its manifest entry, no truncated caption stubs, the lightbox, the theme toggle, and zero console errors. ```bash cd e2e && PORT=5180 node mj-sheets.spec.mjs ``` --- ## Related - `lib/mj-feed.js` — the feed gate (parsers + the in-page expressions) - `lib/mj-feed-reconcile.js` — re-derive every slot's task id from the live feed and rewrite the metas. Importable (`require.main === module` guards the CLI), so the file-IO half is unit-testable without a browser; `npm run test:mj-feed-reconcile` covers it. - `lib/mj-web.js` — the web backend that submits prompts - `docs/MJ-WATCHER.md` — the drop-a-file watcher around all of it