aesthetic.computer / core
Monorepo for Aesthetic.Computer aesthetic.computer
4
fork

Configure Feed

Select the types of activity you want to include in your feed.

core / papers / jeffrey-platter /
at main 3 files
README.md
manifest.json
sync.mjs
README.md

Jeffrey Platter#

An index of canonical photographic references for Jeffrey Alan Scudder — face, body, hands — across the repo, the assets CDN, the vault, and the social silos. A sub-platter within the papers platter, parallel to whistlegraph-platter.

The platter exists so that any pipeline (CV layout, image generation, press kit, lecture slide, video composite) can resolve "a photograph of jeffrey" to a concrete URL or file path with known POIs, focal points, and provenance — instead of reaching for whatever's nearby.

This is index-only: pointers to where canonical images live. Masters live on the assets CDN and in the vault. No binary duplication into papers/.

1. In this repo#

Pieces (capture surfaces)#

These are the AC pieces that produce new photographic frames of jeffrey when he uses them:

POI manifest (single source of truth)#

  • manifest.json — POI manifest for both CDN buckets (8 headshots + 38 candids). Each item carries focal: [x%, y%], pois: [{t: f|b|h, box}], and aspect. POIs originally detected via OpenCV DNN + Haar cascades.
  • sync.mjs — copies manifest.json to consumer-served paths. Run after editing the manifest. Currently writes one target: system/public/give.aesthetic.computer/jeffreys-manifest.json.

Consumers (where canonical jeffrey-images are already used)#

  • system/public/justanothersystem.org/cv.html — bio portrait slideshow with Ken Burns; pulls jeffery-av--07/06/10.jpg from the shoot/ bucket. OG/Twitter card image is jeffery-av--07.jpg. Hardcoded URLs; doesn't read the manifest yet (TBD).
  • system/public/give.aesthetic.computer/index.html — full Jeffreys Ken Burns canvas slideshow. Fetches ./jeffreys-manifest.json at runtime via loadJeffreysManifest() and merges into the slideshow's image index (buildImageIndex()). Originally inlined the same data as JS object literals; lifted to JSON 2026-04-28.
  • recap/ — recap pipeline uses the jeffrey-pvc Professional Voice Clone via /api/say. Audio analogue of canonical-jeffrey: not visual but lives in the same canonical-self category. See recap/SCORE.md.

Adjacent textual / video material#

2. External (assets CDN + sites)#

Hosted dashboard — papers.aesthetic.computer/platter/jeffrey/#

Public sortable/filterable gallery of the 457 confirmed-jeffrey IG selfies from @whistlegraph (face-matched against the AV shoot, described per-image by GPT-4o). Lives under the new /platter/ namespace alongside future per-platter dashboards:

Page is public — the source images were already public on instagram.com/whistlegraph/, and the GPT-4o descriptive layer is treated as part of the public AC research record.

assets.aesthetic.computer/jeffreys/#

Hosted on Digital Ocean Spaces. Sync via npm run assets:sync:down / npm run assets:sync:up (see CLAUDE.md).

jeffreys/shoot/ — Professional AV photoshoot (face-focused headshots)#

55 headshots named jeffery-av--01.jpg through jeffery-av--55.jpg, uploaded 2026-01-05. Bucket audited 2026-04-28 via aws s3 ls; all 55 are present and cataloged in manifest.json under buckets.shoot.items.

Resolution tiers (sizes vary 10x within the sequence — uploads were sized for different consumers):

Tier Range Size Use
master --01..--10 20–26 MB Training crops, print, archival
mid --11..--35 1.7–5.3 MB Slideshows, web reference
web --36..--55 220–340 KB Thumbnails / smallest backgrounds

Per-tier annotation lives in each item's tier and size fields. All headshots share uniform POI framing: aspect: 0.667, focal: [50, 35], pois: [{t:"f", box:[30,15,40,40]}] — face-centered portrait. The framing is a placeholder; a re-run of OpenCV face detection on master-tier frames would yield true bounding boxes.

Note the spelling: filenames use jeffery- (one r), not jeffrey-.

Currently consumed:

  • give.aesthetic.computer Ken Burns slideshow — re-enabled 2026-04-28; merges all 55 into allImagesData (was previously commented out — only 8 cataloged then)
  • justanothersystem.org cv.html — hardcoded URLs for --06/07/10 (OG card uses --07)

jeffreys/<NAME>.{heic,HEIC,jpeg,JPEG} — iPhone master files#

38 masters sit at the top of the jeffreys/ prefix (not in a subdirectory), 1:1 with jpg/ candids. These are the iPhone-original HEIC/JPEG files — the JPGs in candids/ are derivatives. Cataloged in manifest.json under buckets.masters.items with candid_key cross-references and per-file size.

Use this bucket for training crops — masters carry roughly 2× the byte-budget of their JPG derivatives and preserve color depth that JPEG re-encoding lossily compresses. Don't surface masters in slideshows (HEIC has spotty browser support); use the candids/ JPGs for those.

Mixed extensions in the bucket: 32 are heic/HEIC, 6 are jpeg/JPEG. Manifest keys preserve the actual extension on the CDN.

jeffreys/jpg/ — Candids (JPG derivatives of the masters)#

38 enumerated candids (IMG_NNNN.jpg + FullSizeRender) with hand-tuned focal points and POI bounding boxes. Full data lives in manifest.json under buckets.candids.items. Includes:

FullSizeRender, IMG_0260, IMG_0675, IMG_0686, IMG_0688, IMG_0798, IMG_1111, IMG_1577, IMG_1616, IMG_1737, IMG_1809, IMG_2124, IMG_2208, IMG_2280, IMG_2498, IMG_2630, IMG_2658, IMG_2668, IMG_2905, IMG_2913, IMG_3017, IMG_3234, IMG_4281, IMG_4312, IMG_4606, IMG_4894, IMG_4997, IMG_5043, IMG_5050, IMG_5272, IMG_5644, IMG_6342, IMG_6367, IMG_6435, IMG_8080, IMG_8188, IMG_8989, IMG_9795.

Bucket audited 2026-04-28 — exactly 38 entries, manifest is complete, no orphans, no uncataloged files. Each candid has a master field pointing at its HEIC/JPEG counterpart in jeffreys/; prefer the master for training, use the JPG for slideshows and embeds.

POI types: f = face, b = body, h = hand. Aspects vary (0.562 / 0.563 / 0.667 / 0.75 / 0.8 / 1.333). IMG_2124 and IMG_2658 are referenced by name in give.aesthetic.computer/index.html as the splash/background plates.

TBD: locate the script that generated the existing focal/POI values (OpenCV DNN + Haar cascades, per the give-page comment) and check it in to portraits/jeffrey/bin/. The manifest carries the output of that pipeline; the pipeline itself isn't in the repo yet.

screenshots/images/ — First-person POV "selfies" (90 environmental webp captures)#

90 webp screenshots of jeffrey's working environment — laptop screens, hands on the keyboard, desk surfaces, ambient development moments. Captured live while making AC. Treated as canonical first-person POV reference imagery for any pipeline that wants a POV-from-jeffrey look (vs the third-person headshots in shoot/). Mostly portrait orientation (aspect: 0.75), each item has a focal coordinate and pois boxes (t: "s" = scene region).

URL pattern: https://assets.aesthetic.computer/screenshots/images/<name>.webp. Cataloged in manifest.json under buckets.screenshots.items (imported from give.aesthetic.computer/index.html 2026-05-02). Currently consumed by the give Ken Burns slideshow (merged into allImagesData) and available to image-gen pipelines via --refs.

jeffreys/gens/ — Generated images (gpt-image-2 + platter-grounded identity)#

Output bucket for any pipeline that synthesizes a new jeffrey-image conditioned on the platter refs (typically SHOOT_REFS + SELFIE_REFS from portraits/jeffrey/bin/generate-neo.py). One PNG per successful gen, dated, never overwritten.

Filename convention: <context>-<segment>-<YYYY-MM-DDTHHMM>.png — e.g. recap-jeffrey-24h-02_menuband-2026-04-29T1630.png. context says where the gen came from (e.g. recap-jeffrey-24h); segment is whatever locally-meaningful subdivision the producer cares about; the timestamp keeps regens distinct so we never lose a previous take.

Producers:

Manifest entry shape (under buckets.gens.items[<filename>]):

{
  "model": "gpt-image-2",
  "size": "1024x1536",
  "quality": "high",
  "refs": ["portraits/jeffrey/corpus/shoot/jeffery-av--07.jpg", "..."],
  "context": "recap-jeffrey-24h",
  "segment": "02_menuband",
  "generated": "2026-04-29T16:30:00.000Z",
  "bytes": 2520266,
  "prompt": "<full gpt-image-2 prompt — preserved for provenance>"
}

The full prompt is kept inline so a year from now we can answer "what was the metaphor that produced this take?" without cross-referencing the source audience config (which has likely drifted by then).

Push to CDN: npm run assets:sync:up after a batch of gens. The bucket is under the standard system/public/assets/* gitignore — git tracks the manifest entry, not the PNG bytes.

Social silos (canonical public faces)#

Institutional pages carrying jeffrey portraits#

3. Local bulk archive (gitignored)#

Bulk image data lives at portraits/jeffrey/ under gitignored subdirectories — these aren't secrets (vault is for SECRET_FILES only), just bulky regenerable cache:

  • portraits/jeffrey/ig-archive/<account>/ — Instagram bulk dumps. As of 2026-04-29: whistlegraph/ ≈ 4.1 GB, 6,104 jpgs spanning 2014-08-23 → 2026-04-28 (posts + highlights + active stories). Pulled via portraits/jeffrey/bin/ig-archive.fish.
  • portraits/jeffrey/curated/ — outputs of the face-match → vision-describe pipeline. jeffrey-match.jsonl (insightface identity scores), jeffrey-described.jsonl (GPT-4o scene-graph metadata), thumbnails/ (384px), index.html (static browser).
  • TBD: portraits/jeffrey/selfies/ — selfie corpus (raw output from selfie.mjs/snap.mjs/cap.mjs runs); promote the keepers to assets/jeffreys/jpg/ after curation.
  • TBD: portraits/jeffrey/press/ — press headshots and high-res stills sent to institutions.
  • TBD: portraits/jeffrey/shoot-raw/ — uncompressed masters from the AV photoshoot (only web-sized JPGs are on the CDN today).

The only jeffrey-related thing that lives in the vault is the instaloader session cookie at aesthetic-computer-vault/silo/instaloader-sessions/whistlegraph — that one IS a credential (authenticated IG session) and stays vault-side.

4. Instagram archive ingestion (currently BLOCKED)#

The silo Instagram bridge is the path to bulk-importing jeffrey's solo + AC photo archives into the platter.

  • Bridge code: silo/server.mjs on the silo VPS, port 3003
  • Migration report: reports/instagram-api-migration-2026-03-29.md
  • Status (2026-03-29): login rejected on instagram-private-api, @i7m/instagram-cli, and subzeroid/instagrapi-rest. Errors range from IgLoginBadPasswordError to "IP added to the blacklist". Migration to instagrapi 2.3.0 (Python, actively maintained) sketched but not landed.
  • Block resolution path: bypassed 2026-04-29 — instaloader cookie-import (Chrome) succeeded; bulk archive of @whistlegraph (4.1 GB, 6,104 frames, 2014–2026) lives at portraits/jeffrey/ig-archive/whistlegraph/. Silo bridge migration is independent and still pending — only needed if the silo dashboard needs live IG profile/feed queries.

When ingestion comes online, image filenames should map IG shortcodes to date-prefixed names (YYYY-MM-DD-<shortcode>.jpg) so timeline queries are easy. POI re-detection is a separate post-ingest pass.

5. For canonical image generation (the goal)#

The motivating use case for this platter is to feed image-generation pipelines a stable, well-annotated reference set of jeffrey so generated portraits stay on-model.

Pipeline location: portraits/jeffrey/ at the monorepo root. It consumes manifest.json and the CDN buckets; it is the agreed home for training scripts, LoRA artifacts, IP-Adapter ref sets, and any sample outputs.

Currently in place

  • A two-bucket archive (shoot/ headshots + jpg/ candids) hosted on the CDN
  • Per-image POI metadata (face/body/hand bounding boxes + focal point + aspect) for 46 images in manifest.json. Lifted out of give.aesthetic.computer/index.html 2026-04-28.
  • Voice analogue (jeffrey-pvc PVC) already wired into /api/say — establishes the pattern of canonical-self assets being callable as a service

Gaps to close

  • TBD: agree on what "canonical" means for generation:
    • Face LoRA / DreamBooth training corpus — needs ≥20 high-quality face crops with diverse lighting/angles. The 8 shoot/ headshots are too uniform on their own (all focal: [50, 35]); needs candids mixed in.
    • IP-Adapter / FaceID reference set — needs a smaller curated set of 5–10 very clean references.
    • Identity-preserving prompt suffix — needs a stable text description of jeffrey's appearance, derived from photos. Adjacent to but distinct from voice clone metadata.
  • TBD: licensing & consent rails — even though jeffrey is the subject and operator, downstream pipelines that publish generated portraits should have an explicit allowlist of where they're allowed to render him.

Adjacent infrastructure to study

  • arxiv-penrose/ — Penrose pipeline for AC illustrations. Not for portraits, but the "diagrams from data" pattern (declarative spec → rendered image) is a useful analogue.
  • recap/bin/scout.mjs — content-query resolver. Could be extended with a { jeffrey: "shoot" | "candid" | "any", count: N } query that returns N canonical jeffrey URLs for use in slide compositions.

6. Timeline (rough spine)#

  • 2026-01-03 to 2026-01-05 — candids and HEIC masters uploaded to jeffreys/ and jeffreys/jpg/ (38 files each); AV shoot uploaded to jeffreys/shoot/ (55 frames) on 2026-01-05
  • 2026-01–04 — POI manifest authored by hand inside give.aesthetic.computer/index.html; consumed by cv.html (justanothersystem.org) and the Ken Burns slideshow
  • 2026-03-29 — silo IG migration attempted, blocked at login layer (report)
  • 2026-04-28 — platter index created; POI manifest lifted from give.aesthetic.computer/index.html to standalone manifest.json; give-page now fetches the synced copy at runtime; bucket audit added masters/ bucket and tier/size/master annotations; headshots merge re-enabled in give-page; portraits/jeffrey/ scaffolded; silo instagrapi.service staged
  • TBD: candid corpus origin — IMG_NNNN.jpg filenames are iPhone-default; the masters carry EXIF dates that should be parsed and added to the manifest as taken_at

How to extend this index#

  1. When new material appears (a new shoot, a new selfie session keeper, a press photo placement, an ingested IG batch) — add a line under the matching section with a link or asset path.
  2. Use TBD: prefix for known-missing items so they're greppable.
  3. Don't check binaries into papers/. Masters live on the CDN or in the vault; this file points at them.
  4. When the POI manifest gets extracted from give.aesthetic.computer/index.html into a standalone JSON, link it here and update §2 to reflect the new source of truth.
  5. If a sibling manifest.json or pois/ directory grows, note it in §1 (consumers) so other pipelines can find it.