shoots embeddings

shoots embeddings <path> [options]

---

Why this is separate from rate#

If embeddings inherited a preset's bias, anything you trained on it would learn that preset, not you. So it emits the raw signal and lets the trainer decide.

The model name is recorded in the dataset to pin the embedding space — a learned profile can then refuse to be applied to a mismatched backend.

---

Arguments#

Options#

---

Two output modes#

--json — dataset on stdout#

The consolidated dataset, embeddings included, no previews. Good when your consumer already knows how to display the source images (i.e. they are JPEGs).

shoots embeddings ./jpegs --json > dataset.json

--out <dir> — a self-contained bundle#

bundle/
├── embeddings.json
└── previews/
    ├── 000_IMG_0001.jpg
    ├── 001_IMG_0002.jpg
    └── ...

RAW originals are not browser-viewable, so the bundle carries JPEG previews for duel UIs and any web tool. Each result gains a preview field holding a path relative to the dataset file, so the bundle is movable.

Previews come from the embedded RAW preview (via exiftool), resized and EXIF-oriented with sharp.

--previews modes#

---

Examples#

Bundle a RAW shoot for the duel UI#

shoots embeddings D:/Shoots/2026/street-june --out ./bundle

Wrote bundle to ./bundle: embeddings.json + 1204 previews in previews/ (dim 512)

Everything previewed, smaller and lighter#

shoots embeddings ./mixed-catalog --out ./bundle \
  --previews always --preview-size 800 --preview-quality 75

JSON only, no previews#

shoots embeddings ./jpegs --json > dataset.json
shoots embeddings ./raw --out ./bundle --previews never

Human-readable check#

shoots embeddings ./sample

IMG_0001.CR3  dim=512  aspects=7  seed=0.573  [street, urban, candid]
IMG_0002.CR3  dim=512  aspects=7  seed=0.541  [portrait, indoor]

2/2 embedded with clip-vit-b32-int8 (dim 512)

Feed the preference-learning pipeline#

shoots embeddings ~/Pictures/my-best-work --out ./bundle
shoots match import --data ./bundle/embeddings.json --name my-eye
shoots match serve --name my-eye                  # duel at http://127.0.0.1:4576
shoots match train --name my-eye                  # → ~/.shoots/profiles/my-eye.json
shoots rate ./new-shoot --profile my-eye --write-xmp

Full walkthrough: Preference learning.

---

Dataset format#

{
  "command": "embeddings",
  "model": "clip-vit-b32-int8",
  "dim": 512,
  "results": [
    {
      "file": "D:/raw/IMG_0001.CR3",
      "embedding": [0.021374, -0.045912, "… 512 floats …"],
      "aspects": [
        { "name": "overall", "score": 0.612 },
        { "name": "composition", "score": 0.554 }
      ],
      "keywords": ["street", "urban", "candid"],
      "focus": 0.812,
      "aestheticSeed": 0.573,
      "preview": "previews/000_IMG_0001.jpg"
    }
  ],
  "errors": [],
  "summary": { "total": 1204, "embedded": 1204, "failed": 0 }
}

aestheticSeed exists to give an active-learning pairing strategy somewhere to start before any duels exist. Do not mistake it for a rating.

In bundle mode, the --json document on stdout omits results (they are already in the file) and adds out and previews fields instead.

---

Exit codes#

---

See also#

• Preference learning — the full duels → profile pipeline
• rate — the opinionated consumer of the same model
• Rating profiles — what a learned profile looks like