Rating profiles

shoots rate ./raw --profile wedding

---

What a profile decides#

1. WHAT matters — meritWeights#

How the per-aspect CLIP scores combine into a single aesthetic merit. Aspects absent from the map count zero.

2. HOW strict — the star cut-offs and the focus gate#

Where the 0–5 boundaries sit, and how hard technically-soft frames are punished.

The skill-level axis#

For a beginner, technical aspects (exposure, sharpness, composition) should earn stars — nailing them is a real achievement. For a professional they are table stakes, so pro-oriented profiles zero them out and let content decide.

That axis lives in the same two fields: different weights, different thresholds.

---

The aspects#

Seven contrastive aspect pairs, scored zero-shot by CLIP:

Every aspect is always scored and recorded in the sidecar, regardless of the profile's weights. The profile only decides how they aggregate — so you can re-derive a rating under a different profile without re-running inference.

---

Built-in profiles#

street — the default#

Street / documentary — content over craft, unforgiving bar (calibrated)

Technical competence is assumed, so exposure, sharpness and composition are zeroed entirely. Deliberately unforgiving: on a real shoot the mass of frames lands at 0 stars. That is the intended behaviour — you want the tail, not a flattering distribution.

This is the only profile calibrated against a real, hand-judged shoot.

generic#

All-round, forgiving — technical competence counts (prior)

A sensible default for a beginner or a mixed set. Technical craft counts toward the score and the bar is much lower, so a clean, well-made frame already earns a star or two.

portrait#

Portrait — subject & light lead, eyes must be sharp (prior)

The subject and the light on it carry the frame. The focus gate is stricter — a soft portrait is a miss.

wildlife#

Wildlife — sharp subject & behaviour, strict focus (prior)

The hardest focus gate of any profile: a soft animal is simply a miss.

wedding#

Wedding — forgiving, a clean frame already counts (prior)

Deliberately forgiving: on a wedding, a clean well-exposed frame is already a usable pick, and you deliver volume.

Known limit. True emotion and expression scoring is what a wedding profile most wants, and the current model has no such aspect. It arrives with a future model archive. Until then this profile is a reasonable prior, not a calibrated judge.

---

Choosing a profile#

The presets are priors, not truth. Only street is calibrated. The honest way to choose:

# Take 50 frames you have already judged yourself, then compare
for p in street generic portrait wildlife wedding; do
  echo "── $p"
  shoots rate ./sample-50 --profile "$p" --dry-run --json \
    | jq -r '.results[] | "\(.stars)★ \(.file | split("/") | last)"' \
    | sort -r
done

Pick the one that agrees with you most often. If none does — and that is a common, legitimate outcome — train your own.

---

Learned profiles#

The real answer to "none of the presets is my eye" is to learn one from your own judgements. See Preference learning for the full pipeline.

Installing one#

Drop the profile JSON into ~/.shoots/profiles/ and it becomes selectable by filename:

cp my-eye.json ~/.shoots/profiles/
shoots rate ./raw --profile my-eye

Built-in names win over user profiles, so do not name a learned profile street — it would be silently shadowed.

The linear-embedding shape#

Learned profiles use type: "linear-embedding": the merit is a linear head over the CLIP embedding, s(x) = w·x + b, squashed into [0,1]. The star cut-offs then apply to that normalized score exactly as with a built-in profile.

{
  "type": "linear-embedding",
  "name": "my-eye",
  "description": "Learned from 800 duels, June 2026",
  "embeddingModel": "clip-vit-b32-int8",   // guards the embedding space
  "dim": 512,
  "weights": [ /* 512 floats */ ],
  "bias": -0.14,
  "scoreNormalization": { "mean": 0.02, "std": 0.31 },
  "focusReject": 0.3,
  "focusSoft": 0.55,
  "focusSoftCap": 1,
  "aestheticStars": [
    { "min": 0.78, "stars": 5 },
    { "min": 0.62, "stars": 4 },
    { "min": 0.45, "stars": 3 },
    { "min": 0.28, "stars": 2 },
    { "min": 0.12, "stars": 1 }
  ]
}

Validation#

Because these are external JSON files, every field is validated before use. A malformed profile fails loudly with exit 2 rather than producing silent garbage stars:

error: invalid profile my-eye.json: weights must be 512 finite numbers

embeddingModel guards that the profile is applied to the same CLIP space it was learned on. Applying a profile trained on one model archive to another is a category error and is rejected.

Listing what is available#

An unknown profile name prints the full list — built-ins plus everything in ~/.shoots/profiles:

error: unknown rating profile 'foo' (available: street, generic, portrait, wildlife, wedding, my-eye, client-work)

---

Profiles in the pipeline config#

- type: rate
  profile: wedding
  output: xmp

See Pipelines.

---

See also#

• rate — the command that consumes profiles
• Preference learning — train your own
• embeddings — the profile-neutral export profiles are learned from