Merge branch 'dev'

Documents object mode removal cleanup, dead embedding path removal,
FutureWarning fix, variant pyproject sync, and docs/wiki updates.

.github/workflows/docker-publish.yml

@@ -36,15 +36,13 @@ env:
 
 jobs:
   build:
-    name: Build (${{ matrix.platform }})
+    name: Build (linux/amd64)
-    runs-on: ${{ matrix.runner }}
+    runs-on: ubuntu-latest
     strategy:
       matrix:
         include:
           - platform: linux/amd64
             runner: ubuntu-latest
-          - platform: linux/arm64
-            runner: ubuntu-latest
     permissions:
       contents: read
       packages: write
@@ -63,10 +61,6 @@ jobs:
         with:
           ref: ${{ inputs.tag || github.ref }}
 
-      - name: Set up QEMU
-        if: matrix.platform == 'linux/arm64'
-        uses: docker/setup-qemu-action@v4
-
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v4
 
@@ -108,7 +102,7 @@ jobs:
       - name: Upload digest
         uses: actions/upload-artifact@v4
         with:
-          name: digest-${{ matrix.platform == 'linux/amd64' && 'amd64' || 'arm64' }}
+          name: digest-amd64
           path: /tmp/digests/*
           if-no-files-found: error
           retention-days: 1

CHANGELOG.md

@@ -7,6 +7,92 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.4.11] - 2026-06-14
+
+### Removed
+
+- **Object mode pipeline fully removed**: YOLO object detection, SigLIP image classification, `TRAINING_MODE`, and `OBJECT_CLASS` env vars are gone. Frigate has no training API for objects; the ~2 GB model stack (torch, torchvision, transformers, ultralytics) was dead weight.
+- **Dead Immich embedding path removed**: `FaceData.embedding` field and the `immich_embedding` parameter to `get_embedding()` were never consumed by any caller. Both removed along with the NumPy import in `immich_api.py` that existed solely for that path.
+- **Dead `mode` config key removed**: `"mode": "face"` was written into job config dicts in `jobs.py` but never read after object mode removal.
+
+### Fixed
+
+- **InsightFace `FutureWarning` suppressed in crop-alignment path**: the `insightface_app.get()` call in `image_processing.py` now wraps the same `warnings.catch_warnings()` suppressor already present in `embeddings.py`, preventing scikit-image deprecation noise in logs.
+
+### Changed
+
+- **Variant pyproject files synced to current state**: `pyproject-rocm.toml`, `pyproject-cpu.toml`, `pyproject-intel.toml` were at v0.2.13 and still listed torch/transformers/ultralytics. Updated to v0.4.11 and cleaned to face-only deps. Note: corresponding lockfiles (uv-rocm.lock, uv-cpu.lock, uv-intel.lock) need regeneration in their respective platform environments.
+- **`MERGE_DUPLICATE_PEOPLE` documented**: README and wiki now explain the default warn-and-skip behaviour vs. setting `true` for a permanent Immich merge, with irreversibility callout.
+- **Wiki fully updated**: all five wiki pages rewritten to remove object mode references, correct model size (~300 MB InsightFace vs former ~1–2 GB HuggingFace+InsightFace), fix default values (`MAX_AUTO_IMAGES` 80→20, `MIN_FACE_COUNT` 0→3), add `MERGE_DUPLICATE_PEOPLE` coverage, and update GPU verification commands for current ONNX provider API.
+
+## [0.4.10] - 2026-06-14
+
+### Changed
+
+- **`MAX_AUTO_IMAGES` default lowered from `80` to `20`**: winnow is designed to fill the gap where manual Frigate training images don't exist — not to be the primary dataset. A conservative default ensures winnow-imported images remain secondary to hand-picked ones where both exist.
+
+## [0.4.9] - 2026-06-14
+
+### Changed
+
+- **`FRIGATE_SCORE_CEILING` is now dynamic by default**: previously defaulted to `0` (disabled). Now unset (default) enables a self-calibrating novelty gate — below-cap candidates are skipped if their pre-upload Frigate score exceeds the most-redundant tracked file's score. This catches conditions already covered by manually-added Frigate images that winnow cannot track. Set `FRIGATE_SCORE_CEILING=0` to disable entirely; set a positive value (e.g. `0.85`) for a fixed hard ceiling.
+- **Quality replacement branches consolidated**: the Frigate-score and blur-score replacement paths in the upload loop shared identical structure. Merged into a single code path parameterised by score source and comparison direction.
+- **`MIN_FACE_COUNT` default raised from `0` to `3`**: people with fewer than 3 tagged photos produce degenerate training sets; skipping them by default avoids noisy runs.
+- **`STRATEGY=adaptive`** is the new primary name for embedding-based diversity selection; `auto` remains a silent alias for backwards compatibility.
+- **`MERGE_DUPLICATE_PEOPLE` and `TRACE_CROP_SIZE`** added to the README env var table (were in the codebase but undocumented).
+- **CUDA version corrected** in the image tags table (was 13.3, actual base image is 12.8.1).
+
+## [0.4.8] - 2026-06-14
+
+### Changed
+
+- **Tracker write-through cache**: `upload_tracker` now keeps an in-memory copy of each JSON file keyed by its resolved path. All reads after the first hit the cache instead of disk; writes go to both disk and cache atomically. Cuts per-person disk I/O in the upload loop from ~90 reads to ~1, with no API or behaviour changes.
+
+## [0.4.7] - 2026-06-14
+
+### Changed
+
+- **`_dedup_embeddings` pre-allocated buffer**: replaced the grow-on-keep `np.vstack` pattern with a pre-allocated `(Q, D)` buffer filled row-by-row. Eliminates O(K²) copy work and the GC pressure from K intermediate heap allocations while keeping identical arithmetic for the similarity checks.
+- **`_kmedoids` cost computation vectorized**: the Python-level `sum(dist_matrix[i, medoids[labels[i]]] for i in range(n))` generator (called once per swap evaluation) is replaced with `dist_matrix[np.arange(n), np.array(medoids)[labels]].sum()` — a single numpy fancy-index + reduction, ~20–50× faster in the swap loop.
+- **`_reconcile_frigate_mappings` single-write batch**: previously called `record_frigate_file` once per uploaded file, each doing a full JSON load + save (O(L) disk round-trips per person). Now builds the full `{frigate_filename: asset_id}` mapping dict and writes it in one `record_frigate_files_batch` call (O(1) disk round-trip).
+
+## [0.4.6] - 2026-06-14
+
+### Fixed
+
+- **OOM when Immich returns many pages per person**: `fetch_all_assets` now stops fetching once 5000 assets have been collected — the diversity selection pool is already capped at 3000 items, so fetching up to 1,000,000 was wasteful and could exhaust memory on large libraries. 5000 provides ample headroom for the pool cap while bounding per-person memory to ~2 MB.
+- **Non-dict items in Immich asset pages silently skipped**: a malformed or partially-null Immich response page could include `null` or non-object items in the assets array. These are now filtered at fetch time rather than causing `AttributeError` downstream.
+
+## [0.4.5] - 2026-06-14
+
+### Fixed
+
+- **Near-duplicate dedup O(N²) allocation**: `np.vstack(kept_normed)` was rebuilt on every loop iteration even for candidates that would be dropped; the stack is now rebuilt only when a new item is kept, reducing memory pressure significantly for large pools.
+- **`quality_score` falsy-zero in dedup sort**: the sort key used `c.get("quality_score") or 0.0`, which treated a legitimate `quality_score=0.0` identically to a missing key. Changed to an explicit `None` check so zero is preserved as-is, and object-mode candidates (which have no `quality_score`) continue to sort stably to the back.
+- **Post-dedup pool not re-checked against limit**: after near-duplicate removal the pool could silently shrink below the requested limit with no warning. A second `len < limit` guard now fires after dedup and emits the same "Only N embeddings" warning that the pre-dedup guard does.
+- **`mark_rejected` could miss plain-text 400 bodies longer than 100 bytes**: `error_detail = resp.text[:100]` was being searched for the keyword `"face"` to gate `mark_rejected()`, so a response body with `"face"` after byte 100 would never mark the asset rejected and it would be retried on every future run. The `"face"` check now uses the full response body; truncation is kept only for the displayed snippet.
+- **`_safe_person_dir` raised ValueError for all person names when `output_dir` resolved to `/`**: `base + os.sep` produced `"//"` when base was `"/"`, and valid paths like `/alice` don't start with `"//"`. Fixed by using `base` directly as the prefix when `base == os.sep`.
+
+## [0.4.4] - 2026-06-14
+
+### Added
+
+- **`RESET_PERSON=*` bulk reset**: resets every tracked person at once (deletes their Frigate training files and clears tracker data). Any other value still resets that specific person by name. If a person is literally named `*` they are reset as part of the bulk operation, and a warning is printed to clarify this.
+- **Near-duplicate removal before diversity selection**: a greedy dedup pass now runs after embedding collection and before clustering. Candidates within 0.20 cosine distance of a higher-quality image are dropped, eliminating burst shots and same-event lookalike photos that produce redundant training images. The best-quality frame from each near-identical group is kept. Dropped count is logged per person.
+
+### Fixed
+
+- **HTTP 500 upload errors no longer show Frigate's misleading "Try restarting Frigate" message**: the response body is now logged at debug level only. HTTP 400 detail (e.g. "No face was detected") is still shown since it is actionable.
+- **`RuntimeWarning: Mean of empty slice`** when a person has only one image after quality filtering: `_compute_adaptive_threshold` now returns the floor value immediately when there are no pairwise distances to sample, and the k-medoids cluster count is floored at 1 to prevent `k=0`.
+- **Path traversal guard on output directory**: person names with `../` sequences or absolute paths (e.g. `/etc`) are now rejected before any filesystem operation, logging an error and skipping the job rather than writing outside the output tree.
+
+## [0.4.3] - 2026-06-14
+
+### Added
+
+- **InsightFace landmark-based face crop alignment**: face crops for Frigate training are now aligned using InsightFace's `norm_crop` (ArcFace 112×112 alignment with 5-point facial landmarks). Previously, Immich's API returned only bounding boxes with no landmarks, so `align_face()` was dead code and crops were plain bbox slices — resulting in misaligned or partial crops (e.g. foreheads). The fix runs InsightFace detection on an expanded region around the Immich bbox, finds the nearest face, and uses its keypoints for proper alignment. Controlled by `ENABLE_FACE_ALIGNMENT` (default `true`).
+- **Duplicate Immich person detection and handling**: when multiple Immich person records share the same name, winnow now detects this at startup and warns with a per-group summary. Without handling, two jobs would run for the same Frigate folder and overwrite each other's output. By default (`MERGE_DUPLICATE_PEOPLE=false`) only the first person per name is processed. Set `MERGE_DUPLICATE_PEOPLE=true` to permanently merge duplicate records inside Immich (keeps the person with the most assets).
+
 ## [0.4.2] - 2026-06-13
 
 ### Changed

CONTRIBUTING.md

@@ -36,6 +36,10 @@ CI runs both on every push and PR to `main` and `dev`. PRs must pass before merg
 - Keep the `CHANGELOG.md` entry in the `[Unreleased]` section updated.
 - Commit messages should be plain English describing what changed and why.
 
+## Development Tooling
+
+Development uses Claude Code (Anthropic) for implementation assistance. All code is reviewed and the final call on design, behavior, and what ships is made by the maintainer. Contributions from humans are equally welcome.
+
 ## License
 
 By submitting a contribution you agree that your work will be released under the project's [AGPLv3+ license](LICENSE).

Dockerfile

@@ -67,7 +67,7 @@ FROM base-${TARGETARCH}-${VARIANT} AS runtime
 ARG VARIANT=gpu
 ARG VERSION=dev
 LABEL org.opencontainers.image.title="winnow" \
-      org.opencontainers.image.description="Selects diverse, high-quality photos from Immich as training data for Frigate face recognition and object classification." \
+      org.opencontainers.image.description="Selects diverse, high-quality photos from Immich as training data for Frigate face recognition." \
       org.opencontainers.image.source="https://github.com/sudolulo/winnow" \
       org.opencontainers.image.licenses="AGPL-3.0-or-later" \
       org.opencontainers.image.version="${VERSION}"
@@ -116,7 +116,7 @@ https://repositories.intel.com/graphics/ubuntu jammy flex" \
     fi
 
 RUN groupadd -g 568 apps && useradd -u 568 -g apps -m -s /bin/bash appuser \
-    && mkdir -p /models/.insightface /models/huggingface \
+    && mkdir -p /models/.insightface \
     && chown -R appuser:apps /app /models
 
 WORKDIR /app
@@ -124,7 +124,7 @@ USER appuser
 # PYTHONPATH=/app makes the winnow package importable from the entry point script.
 # uv sync builds the wheel before winnow/ is COPY'd, so site-packages has only
 # the dist-info. Explicitly adding /app lets Python find winnow/__init__.py there.
-ENV HF_HOME=/models/huggingface INSIGHTFACE_HOME=/models/.insightface PYTHONPATH=/app
+ENV INSIGHTFACE_HOME=/models/.insightface PYTHONPATH=/app
 
 HEALTHCHECK CMD test -f /app/entrypoint.sh || exit 1
 ENTRYPOINT ["tini", "--", "/app/entrypoint.sh"]

README.md

@@ -2,16 +2,18 @@
 
 [![Docker](https://github.com/sudolulo/winnow/actions/workflows/docker-publish.yml/badge.svg)](https://github.com/sudolulo/winnow/actions/workflows/docker-publish.yml) [![Test](https://github.com/sudolulo/winnow/actions/workflows/test.yml/badge.svg)](https://github.com/sudolulo/winnow/actions/workflows/test.yml) [![GitHub release](https://img.shields.io/github/v/release/sudolulo/winnow)](https://github.com/sudolulo/winnow/releases/latest) [![License: AGPL v3](https://img.shields.io/badge/License-AGPL_v3-blue.svg)](LICENSE) [![Immich](https://img.shields.io/badge/Immich-v1.106%2B-blueviolet)](https://immich.app) [![Frigate](https://img.shields.io/badge/Frigate-Ready-brightgreen)](https://frigate.video)
 
+> **Note:** winnow's approach to training Frigate face recognition is not an officially documented workflow — results may vary.
+
 > **Early Development — Use With Caution**
 > winnow is functional but still maturing. Features that modify your Frigate training data — quality replacement, stale mapping cleanup — can remove images from your dataset and are not yet battle-tested at scale. Review the logs after each run and keep backups of your Frigate face training directory until you are confident in the results.
 
 **Docs:** [Setup](https://github.com/sudolulo/winnow/wiki/Setup) · [Troubleshooting](https://github.com/sudolulo/winnow/wiki/Troubleshooting) · [FAQ](https://github.com/sudolulo/winnow/wiki/FAQ)
 
-`winnow` pulls photos from your [Immich](https://immich.app) library, selects the most diverse and highest-quality subset using AI embeddings, and delivers them as training data for [Frigate](https://frigate.video)'s face recognition and object classification models.
+`winnow` pulls photos from your [Immich](https://immich.app) library, selects the most diverse and highest-quality subset using AI embeddings, and delivers them as training data for [Frigate](https://frigate.video)'s face recognition.
 
-Frigate's face recognition is only as good as its training data — and the key quality metric is **diversity**, not volume. A hundred photos from the same week teach the model one lighting condition. What you need is a spread: different years, different angles, different lighting, different contexts. Your photo library already has that data. winnow finds and delivers the right subset automatically.
+The best Frigate training data is images you curate manually — photos taken specifically for recognition, in controlled conditions, uploaded directly through Frigate's UI. For people you can do that for, do it. winnow is for everyone else: people in your library you want Frigate to recognise but don't have dedicated training photos for. It mines your existing Immich library for the most diverse spread of real-world appearances and fills the gap.
 
-> **winnow only touches files it uploaded.** Faces added to Frigate manually through its UI are never deleted, replaced, or modified — not by quality replacement, not by `RESET_PERSON`, not by stale cleanup. If you have a curated training set you want to keep, it is safe.
+> **winnow only touches files it uploaded.** Faces added to Frigate manually through its UI are never deleted, replaced, or modified — not by quality replacement, not by `RESET_PERSON`, not by stale cleanup. Your manually curated images are always the primary dataset; winnow only adds to it.
 
 ---
 
@@ -37,49 +39,46 @@ Immich library
       │
       ▼
 4. Compute embeddings from the same preview thumbnails
-   • Faces   → InsightFace (ArcFace / Buffalo_L)  → 512-dim vector
+   • InsightFace (ArcFace / Buffalo_L) → 512-dim vector
-   • Objects → SigLIP (Vision Transformer)         → 768-dim vector
       │
       ▼
-5. Diversity selection
+5. Near-duplicate removal — greedy cosine-distance pass drops burst shots
+   and near-identical photos before clustering runs; the highest-quality
+   image from each near-duplicate group is kept
+      │
+      ▼
+6. Diversity selection
    • K-Medoids clustering → one representative per natural group
    • Farthest Point Sampling → fill remaining slots with maximally spread picks
-   • Hard example weighting — unusual angles and low-confidence detections
+   • Hard example weighting — low-confidence detections get a distance boost
-     are biased toward selection, since those are where models tend to fail
+     so unusual angles and harder looks are preferred over easy frontals
-   • Auto mode: stops when similarity to the existing set exceeds a threshold
+   • Adaptive mode: stops when the next candidate is too similar to those already
-     (20 % of median pairwise distance for faces, 10 % for objects)
+     selected (distance threshold = 20 % of median pairwise distance for
+     faces, 10 % for objects)
       │
       ▼
-6. Download full-resolution originals from Immich
+7. Download full-resolution originals from Immich
       │
       ▼
-7. Crop and process
+8. Crop and process — EXIF-corrected, landmark-aligned 112×112 crop (ArcFace format)
-   • Face mode: EXIF-corrected, landmark-aligned 112×112 crop (ArcFace format)
-   • Object mode: YOLOv9c detection → one crop per matched instance
       │
       ▼
-8. Deliver
+9. Deliver — upload crops to Frigate's face registration API
-   • Face mode: upload crops to Frigate's face registration API
+   ↳ below MAX_AUTO_IMAGES — upload, unless the novelty gate
-     ↳ below MAX_AUTO_IMAGES — upload freely
+     (FRIGATE_SCORE_CEILING) determines the candidate is already
+     covered by the current training set
    ↳ at cap + QUALITY_REPLACEMENT=true — with Frigate scoring active,
      swap the most redundant tracked image (highest pre-upload recognize
      score) if the candidate is more novel (lower score); falling back to
      blur-score comparison when no Frigate scores are available; manually
      added files are never touched
    ↳ at cap + QUALITY_REPLACEMENT=false — skip this person
-   • Object mode: save crops to disk → place into your Frigate data directory
 ```
 
-Uploaded and rejected asset IDs are persisted across runs. The same image is never processed twice; Frigate rejections are permanently skipped unless `RETRY_REJECTED=true`.
+Uploaded and rejected asset IDs are persisted across runs. The same image is never processed twice; rejected assets are permanently skipped unless `RETRY_REJECTED=true`.
 
 ---
 
-## Modes
-
-**Face mode** (default) — extracts face crops using Immich's bounding box metadata, applies EXIF orientation correction, and aligns them to ArcFace's standard 112×112 format using 5-point facial landmarks. Crops are uploaded directly to Frigate's face registration API.
-
-**Object mode** — runs each full-resolution image through YOLOv9c to detect instances of a target class (dog, cat, car, etc.), crops each detection, and saves it to the output directory. Frigate has no API for uploading object training data; place the crops into your Frigate data directory manually.
-
 ---
 
 ## Running in Docker
@@ -88,7 +87,7 @@ Uploaded and rejected asset IDs are persisted across runs. The same image is nev
 
 | Tag | Arch | Acceleration |
 | :-- | :-- | :-- |
-| `:latest` | amd64 + arm64 | NVIDIA CUDA 13.3 (amd64) · requires [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html) |
+| `:latest` | amd64 + arm64 | NVIDIA CUDA 12.8 (amd64) · requires [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html) |
 | `:rocm` | amd64 | AMD ROCm · pass `/dev/kfd` + `/dev/dri` |
 | `:intel` | amd64 | Intel Arc / iGPU via OpenVINO · pass `/dev/dri`, set `OPENVINO_DEVICE=GPU` |
 | `:cpu` | amd64 + arm64 | CPU only · ~2 GB smaller · no GPU required |
@@ -106,7 +105,7 @@ services:
       - FRIGATE_URL=http://192.168.1.10:5000
       - CRON_SCHEDULE=0 3 * * 0
     volumes:
-      - /path/to/models:/models
+      - /path/to/models:/models    # INSIGHTFACE_HOME — persists Buffalo_L model (~300 MB)
       - /path/to/cache:/app/.if_cache
       - /path/to/output:/app/frigate_train
     deploy:
@@ -152,7 +151,7 @@ See [compose.yml](compose.yml) for the full annotated example with all options.
 | *(empty string)* | Stay alive, run nothing — trigger manually with `docker exec -it winnow winnow` |
 | Cron expression | Run on startup, then repeat on schedule |
 
-In scheduled mode the process (and loaded models) stays resident between runs. The first run after a fresh install downloads the embedding models (~1–2 GB); subsequent runs use the cached models from the mounted volume.
+In scheduled mode the process (and loaded models) stays resident between runs. The first run after a fresh install downloads InsightFace Buffalo_L (~300 MB); subsequent runs use the cached model from the mounted volume.
 
 ---
 
@@ -170,11 +169,9 @@ In scheduled mode the process (and loaded models) stays resident between runs. T
 
 | Variable | Default | Description |
 | :--- | :--- | :--- |
-| `TRAINING_MODE` | `face` | `face` — upload crops to Frigate; `object` — save crops to disk |
+| `STRATEGY` | `adaptive` | `adaptive` — embedding-based diversity selection, stops when candidates become redundant; `standard` — fixed 30 images; `broad` — fixed 100 images |
-| `STRATEGY` | `auto` | `auto` (embedding-based adaptive), `standard` (30 images), `broad` (100 images) |
 | `LIMIT` | *(unset)* | Exact image count — overrides `STRATEGY` |
-| `OBJECT_CLASS` | `dog` | Target class for object mode (any YOLO class: `dog`, `cat`, `car`, etc.) |
+| `AUTO_MODE` | *(auto)* | Skip interactive prompts and process all people unattended — auto-detected when no TTY is present (Docker, cron); set `true` to force in a terminal |
-| `AUTO_MODE` | *(auto)* | Force non-interactive mode in a terminal; auto-detected otherwise |
 | `VERBOSE` | `false` | Enable DEBUG-level console output (log file is always DEBUG) |
 
 ### People Filtering
@@ -183,23 +180,33 @@ In scheduled mode the process (and loaded models) stays resident between runs. T
 | :--- | :--- | :--- |
 | `ONLY_PEOPLE` | *(unset)* | Comma-separated whitelist — process only these people |
 | `SKIP_PEOPLE` | *(unset)* | Comma-separated list — skip these people |
-| `MIN_FACE_COUNT` | `0` | Skip people with fewer than N tagged assets in Immich |
+| `MIN_FACE_COUNT` | `3` | Skip people with fewer than N tagged assets in Immich |
+| `MERGE_DUPLICATE_PEOPLE` | `false` | When Immich has duplicate entries for the same person (same face split across multiple names), merge their asset pools before processing. Without this, each duplicate group emits a warning and is skipped |
 | `YEARS_FILTER` | `10` | Ignore images older than N years |
 
+> **Duplicate people detection** — winnow warns at startup if the same name appears on multiple Immich person records (a common side-effect of Immich's face clustering creating separate pools for the same individual). By default (`false`) it logs the duplicates, keeps only the person with the most assets, and skips the rest — no data is changed. Set `MERGE_DUPLICATE_PEOPLE=true` to permanently merge each duplicate group inside Immich (the person with the most assets absorbs the others). **This modifies Immich and cannot be undone.** Only enable it once you've verified the duplicates are actually the same person.
+
 ### Image Quality
 
 | Variable | Default | Description |
 | :--- | :--- | :--- |
+| `MAX_AUTO_IMAGES` | `20` | Maximum training images per person in Frigate |
+| `QUALITY_REPLACEMENT` | `true` | When at cap, swap a weaker tracked image for a better candidate. With Frigate scoring active, targets the most redundant image (highest pre-upload recognize score); otherwise uses blur score. Never touches manually added Frigate files. Set `false` to skip people at cap |
+
+#### Advanced Tuning *(calibrated — do not adjust)*
+
+These defaults are tuned for Frigate's ArcFace requirements. winnow will warn on launch if any are set. Image quality issues caused by non-default values will not be investigated.
+
+| Variable | Default | Description |
+| :--- | :--- | :--- |
+| `ENABLE_FRIGATE_SCORES` | `true` | Call Frigate's recognize endpoint pre-upload to store diversity scores used for quality replacement. Adds ~200 ms per upload. Disabling also disables the below-cap novelty gate |
+| `FRIGATE_SCORE_CEILING` | *(unset)* | Below-cap novelty gate. Unset: dynamic — skips candidates whose Frigate score exceeds the most-redundant tracked file's score, auto-calibrates each run. `0`: disable entirely. Positive value (e.g. `0.85`): fixed hard ceiling |
 | `MIN_FACE_WIDTH` | `90` | Minimum face crop width in pixels |
 | `FACE_MARGIN` | `0.15` | Padding around bounding box crop (fraction of face size) |
 | `ENABLE_FACE_ALIGNMENT` | `true` | Align to ArcFace 112×112 format using facial landmarks |
 | `USE_FULL_RESOLUTION` | `true` | Download full-resolution originals rather than preview thumbnails |
 | `MIN_CONFIDENCE` | `0.7` | Minimum Immich face detection confidence |
 | `BLUR_THRESHOLD` | `120.0` | Laplacian variance threshold — lower accepts more blur |
-| `MAX_AUTO_IMAGES` | `80` | Maximum training images per person in Frigate |
-| `QUALITY_REPLACEMENT` | `true` | When at cap, swap a weaker tracked image for a better candidate. With Frigate scoring active, targets the most redundant image (highest pre-upload recognize score); otherwise uses blur score. Never touches manually added Frigate files. Set `false` to skip people at cap |
-| `FRIGATE_SCORE_CEILING` | `0.0` | Skip uploads whose pre-upload Frigate recognize score exceeds this value — they are already well-covered. `0` disables; requires at least one prior run to have scores |
-| `ENABLE_FRIGATE_SCORES` | `true` | Call Frigate's recognize endpoint pre-upload to store diversity scores used for quality replacement. Adds ~200 ms per upload. Disable to use blur-score replacement only |
 
 ### GPU & Models
 
@@ -209,22 +216,22 @@ In scheduled mode the process (and loaded models) stays resident between runs. T
 | `OPENVINO_DEVICE` | `CPU` | Intel variant only: set `GPU` to use Arc or iGPU; default runs on CPU |
 | `ENABLE_CACHE` | `true` | Cache computed embeddings to disk (speeds up re-runs on the same library) |
 | `CACHE_DIR` | `.if_cache` | Path for embedding cache and upload tracker files |
-| `HF_HOME` | *(system)* | HuggingFace model cache path (SigLIP) |
 | `INSIGHTFACE_HOME` | *(system)* | InsightFace model cache path (Buffalo_L) |
 
 ### Output
 
 | Variable | Default | Description |
 | :--- | :--- | :--- |
-| `OUTPUT_DIR` | `./frigate_train` | Directory for object-mode crops and the `winnow.log` file. In Docker, set this via the volume mount instead. |
+| `OUTPUT_DIR` | `./frigate_train` | Directory where face crops are staged before upload and where `winnow.log` is written. In Docker, set this via the volume mount instead. |
 
 ### Tracker Overrides *(one-shot — remove after use)*
 
 | Variable | Default | Description |
 | :--- | :--- | :--- |
 | `DRY_RUN` | `false` | Preview selection without downloading or uploading |
-| `RETRY_REJECTED` | `false` | Re-attempt assets previously rejected by Frigate |
+| `RETRY_REJECTED` | `false` | Re-attempt all previously rejected assets (low-confidence skips, Frigate rejections, and other permanent exclusions) |
-| `RESET_PERSON` | *(unset)* | Clear upload history for one person and delete their winnow-managed Frigate training files so the next run starts fresh. Manually added Frigate files are never touched |
+| `RESET_PERSON` | *(unset)* | Set to a person's name to clear their upload history and delete their winnow-managed Frigate training files so the next run starts fresh. Set to `*` to reset all tracked people at once. Manually added Frigate files are never touched |
+| `TRACE_CROP_SIZE` | *(unset)* | Debug: print all tracked crops whose width or height matches this pixel value, then exit |
 
 ### Scheduling
 
@@ -245,14 +252,14 @@ uv run winnow
 
 Requires Python 3.13+ and [uv](https://astral.sh/uv). An NVIDIA, AMD, or Intel GPU is recommended — CPU mode works but embedding computation is slower.
 
-When run with a terminal attached, winnow starts an interactive session: select which people to process and choose a strategy (auto, standard, broad, or a custom count) per person. Without a TTY — Docker, cron, or `AUTO_MODE=true` — it processes all people automatically using the configured defaults.
+When run with a terminal attached, winnow starts an interactive session: select which people to process and choose a strategy (adaptive, standard, broad, or a custom count) per person. Without a TTY — Docker, cron, or `AUTO_MODE=true` — it processes all people unattended using the configured defaults.
 
 ---
 
 ## Requirements
 
 - **Immich** v1.106+
-- **Frigate** v0.16+ (face mode only — object mode has no Frigate dependency)
+- **Frigate** v0.16+
 - **GPU** recommended: NVIDIA (CUDA), AMD (ROCm), or Intel (Arc / iGPU via OpenVINO)
 - **Python** 3.13+
 

compose.yml

@@ -13,19 +13,16 @@ services:
       # Set AUTO_MODE=true to force auto mode in an interactive terminal.
       # To run interactively: docker exec -it winnow winnow
       # - VERBOSE=true                # Enable DEBUG-level console output
-      # TRAINING_MODE: face = upload to Frigate face recognition API
-      #                object = save crops to output dir for manual Frigate placement
-      - TRAINING_MODE=face
       # STRATEGY: auto = objective diversity (recommended), standard = 30 imgs, broad = 100 imgs
       - STRATEGY=auto
       # - LIMIT=50                    # Custom image count; overrides STRATEGY preset
-      # - OBJECT_CLASS=dog            # Object label for object mode (e.g. dog, cat, car)
 
       # ── People Filtering ──────────────────────────────────────────────────
       # - ONLY_PEOPLE=John,Jane       # Comma-separated; process only these people
       # - SKIP_PEOPLE=Unknown         # Comma-separated; skip these people
       # - MIN_FACE_COUNT=5            # Skip people with fewer than N assets in Immich
       # - YEARS_FILTER=10             # Only include images from the last N years (default: 10)
+      # - MERGE_DUPLICATE_PEOPLE=true # Auto-merge Immich people with the same name (keeps most assets)
 
       # ── Image Quality ─────────────────────────────────────────────────────
       # - MIN_FACE_WIDTH=50           # Minimum face width in pixels (default: 50)
@@ -34,14 +31,13 @@ services:
       # - USE_FULL_RESOLUTION=true    # Use full-res images vs thumbnails (default: true)
       # - MIN_CONFIDENCE=0.7          # Minimum face detection confidence (default: 0.7)
       # - BLUR_THRESHOLD=100.0        # Laplacian blur threshold; lower = accept more blur (default: 100.0)
-      # - MAX_AUTO_IMAGES=80          # Hard cap on auto-diversity selection (default: 80)
+      # - MAX_AUTO_IMAGES=80          # Hard cap on auto-diversity selection (default: 20)
 
       # ── Caching & Models ──────────────────────────────────────────────────
       # - FORCE_CPU=true              # Disable GPU, fall back to CPU
       # - OPENVINO_DEVICE=GPU         # Intel variant only: use Arc/iGPU instead of CPU (default: CPU)
       # - ENABLE_CACHE=false          # Disable embedding cache (default: true)
       - CACHE_DIR=/app/.if_cache
-      - HF_HOME=/models/huggingface
       - INSIGHTFACE_HOME=/models/.insightface
 
       # ── Tracker overrides (one-shot, remove after use) ────────────────────

pyproject-cpu.toml

@@ -1,7 +1,7 @@
 [project]
 name = "winnow"
-version = "0.2.13"
+version = "0.4.11"
-description = "Immich to Frigate training sets"
+description = "Selects diverse, high-quality photos from Immich as training data for Frigate face recognition."
 license = "AGPL-3.0-or-later"
 requires-python = ">=3.13"
 authors = [{ name = "Holden Salomon", email = "holden@arch.fyi" }]
@@ -23,10 +23,6 @@ dependencies = [
     "python-dotenv>=1.2.1",
     "requests>=2.32.5",
     "rich>=14.2.0",
-    "torch>=2.12.0",
-    "torchvision>=0.27.0",
-    "transformers>=5.12.0",
-    "ultralytics>=8.4.66",
 ]
 
 [project.scripts]
@@ -34,25 +30,13 @@ winnow = "winnow.cli:main"
 
 [project.urls]
 Repository = "https://github.com/sudolulo/winnow"
+Changelog = "https://github.com/sudolulo/winnow/blob/main/CHANGELOG.md"
 
 [tool.uv]
 required-environments = [
     "sys_platform == 'linux' and platform_machine == 'x86_64'",
 ]
 
-[tool.uv.sources]
-torch = [
-    { index = "pytorch-cpu", marker = "sys_platform == 'linux' and platform_machine == 'x86_64'" },
-]
-torchvision = [
-    { index = "pytorch-cpu", marker = "sys_platform == 'linux' and platform_machine == 'x86_64'" },
-]
-
-[[tool.uv.index]]
-name = "pytorch-cpu"
-url = "https://download.pytorch.org/whl/cpu"
-explicit = true
-
 [dependency-groups]
 dev = [
     "pytest>=8.0",
@@ -81,9 +65,6 @@ numpy = "numpy"
 onnxruntime = "onnxruntime"
 requests = "requests"
 rich = "rich"
-torch = "torch"
-transformers = "transformers"
-ultralytics = "ultralytics"
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]

pyproject-intel.toml

@@ -1,7 +1,7 @@
 [project]
 name = "winnow"
-version = "0.2.13"
+version = "0.4.11"
-description = "Immich to Frigate training sets"
+description = "Selects diverse, high-quality photos from Immich as training data for Frigate face recognition."
 license = "AGPL-3.0-or-later"
 requires-python = ">=3.13"
 authors = [{ name = "Holden Salomon", email = "holden@arch.fyi" }]
@@ -23,10 +23,6 @@ dependencies = [
     "python-dotenv>=1.2.1",
     "requests>=2.32.5",
     "rich>=14.2.0",
-    "torch>=2.12.0",
-    "torchvision>=0.27.0",
-    "transformers>=5.12.0",
-    "ultralytics>=8.4.66",
 ]
 
 [project.scripts]
@@ -34,6 +30,7 @@ winnow = "winnow.cli:main"
 
 [project.urls]
 Repository = "https://github.com/sudolulo/winnow"
+Changelog = "https://github.com/sudolulo/winnow/blob/main/CHANGELOG.md"
 
 [tool.uv]
 conflicts = [
@@ -47,19 +44,6 @@ required-environments = [
     "sys_platform == 'linux' and platform_machine == 'x86_64'",
 ]
 
-[tool.uv.sources]
-torch = [
-    { index = "pytorch-cpu", marker = "sys_platform == 'linux' and platform_machine == 'x86_64'" },
-]
-torchvision = [
-    { index = "pytorch-cpu", marker = "sys_platform == 'linux' and platform_machine == 'x86_64'" },
-]
-
-[[tool.uv.index]]
-name = "pytorch-cpu"
-url = "https://download.pytorch.org/whl/cpu"
-explicit = true
-
 [dependency-groups]
 dev = [
     "pytest>=8.0",
@@ -88,9 +72,6 @@ numpy = "numpy"
 onnxruntime-openvino = "onnxruntime"
 requests = "requests"
 rich = "rich"
-torch = "torch"
-transformers = "transformers"
-ultralytics = "ultralytics"
 
 [tool.deptry.per_rule_ignores]
 DEP002 = ["onnxruntime-openvino"]

pyproject-rocm.toml

@@ -1,7 +1,7 @@
 [project]
 name = "winnow"
-version = "0.2.13"
+version = "0.4.11"
-description = "Immich to Frigate training sets"
+description = "Selects diverse, high-quality photos from Immich as training data for Frigate face recognition."
 license = "AGPL-3.0-or-later"
 requires-python = ">=3.13"
 authors = [{ name = "Holden Salomon", email = "holden@arch.fyi" }]
@@ -23,10 +23,6 @@ dependencies = [
     "python-dotenv>=1.2.1",
     "requests>=2.32.5",
     "rich>=14.2.0",
-    "torch>=2.5.0",
-    "torchvision>=0.20.0",
-    "transformers>=5.12.0",
-    "ultralytics>=8.4.66",
 ]
 
 [project.scripts]
@@ -34,6 +30,7 @@ winnow = "winnow.cli:main"
 
 [project.urls]
 Repository = "https://github.com/sudolulo/winnow"
+Changelog = "https://github.com/sudolulo/winnow/blob/main/CHANGELOG.md"
 
 [tool.uv]
 index-strategy = "unsafe-best-match"
@@ -48,18 +45,6 @@ required-environments = [
     "sys_platform == 'linux' and platform_machine == 'x86_64'",
 ]
 
-[tool.uv.sources]
-torch = [
-    { index = "pytorch-rocm63", marker = "sys_platform == 'linux' and platform_machine == 'x86_64'" },
-]
-torchvision = [
-    { index = "pytorch-rocm63", marker = "sys_platform == 'linux' and platform_machine == 'x86_64'" },
-]
-
-[[tool.uv.index]]
-name = "pytorch-rocm63"
-url = "https://download.pytorch.org/whl/rocm6.3"
-
 [dependency-groups]
 dev = [
     "pytest>=8.0",
@@ -88,9 +73,6 @@ numpy = "numpy"
 onnxruntime-rocm = "onnxruntime"
 requests = "requests"
 rich = "rich"
-torch = "torch"
-transformers = "transformers"
-ultralytics = "ultralytics"
 
 [tool.deptry.per_rule_ignores]
 DEP002 = ["onnxruntime-rocm"]

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "winnow"
-version = "0.4.2"
+version = "0.4.11"
-description = "Selects diverse, high-quality photos from Immich as training data for Frigate face recognition and object classification."
+description = "Selects diverse, high-quality photos from Immich as training data for Frigate face recognition."
 license = "AGPL-3.0-or-later"
 requires-python = ">=3.13"
 authors = [{ name = "Holden Salomon", email = "holden@arch.fyi" }]
@@ -27,10 +27,6 @@ dependencies = [
     "python-dotenv>=1.2.1",
     "requests>=2.32.5",
     "rich>=14.2.0",
-    "torch>=2.12.0",
-    "torchvision>=0.27.0",
-    "transformers>=5.12.0",
-    "ultralytics>=8.4.66",
 ]
 
 [project.scripts]
@@ -53,27 +49,6 @@ required-environments = [
     "sys_platform == 'linux' and platform_machine == 'aarch64'",
 ]
 
-[tool.uv.sources]
-torch = [
-    { index = "pytorch-cu126", marker = "sys_platform == 'linux' and platform_machine == 'x86_64'" },
-    { index = "pytorch-cpu", marker = "sys_platform == 'linux' and platform_machine == 'aarch64'" },
-    { index = "pytorch-cpu", marker = "sys_platform != 'linux'" },
-]
-torchvision = [
-    { index = "pytorch-cu126", marker = "sys_platform == 'linux' and platform_machine == 'x86_64'" },
-    { index = "pytorch-cpu", marker = "sys_platform == 'linux' and platform_machine == 'aarch64'" },
-    { index = "pytorch-cpu", marker = "sys_platform != 'linux'" },
-]
-
-[[tool.uv.index]]
-name = "pytorch-cu126"
-url = "https://download.pytorch.org/whl/cu126"
-explicit = true
-
-[[tool.uv.index]]
-name = "pytorch-cpu"
-url = "https://download.pytorch.org/whl/cpu"
-explicit = true
 
 [dependency-groups]
 dev = [
@@ -104,9 +79,6 @@ nvidia-cudnn-cu12 = "nvidia.cudnn"
 onnxruntime-gpu = "onnxruntime"
 requests = "requests"
 rich = "rich"
-torch = "torch"
-transformers = "transformers"
-ultralytics = "ultralytics"
 
 [tool.deptry.per_rule_ignores]
 DEP002 = ["onnxruntime-gpu", "nvidia-cudnn-cu12"]

scheduler.py

@@ -16,7 +16,6 @@ except ImportError:
 from winnow.cli import main
 
 SCHEDULE = os.environ["CRON_SCHEDULE"]
-MODELS_DIR = os.environ.get("HF_HOME", "/models/huggingface")
 INSIGHTFACE_HOME = os.environ.get("INSIGHTFACE_HOME", "/models/.insightface")
 
 logger = logging.getLogger(__name__)
@@ -24,11 +23,8 @@ logger = logging.getLogger(__name__)
 
 def check_models() -> None:
     buffalo = Path(INSIGHTFACE_HOME) / "models" / "buffalo_l"
-    hf_hub = Path(MODELS_DIR) / "hub"
     if not buffalo.exists():
         print("  InsightFace Buffalo_L not found — will download on first run", flush=True)
-    if not (hf_hub.exists() and any(hf_hub.iterdir())):
-        print("  HuggingFace models not found — will download on first run", flush=True)
 
 
 NOW = time.time()

scripts/benchmark.py

@@ -2,8 +2,8 @@
 """
 winnow inference benchmark: GPU vs CPU throughput.
 
-Measures InsightFace (face mode) and SigLIP (object mode) latency and
+Measures InsightFace (ArcFace) latency and throughput.
-throughput. Run with FORCE_CPU=true for CPU-only baseline.
+Run with FORCE_CPU=true for CPU-only baseline.
 
 Usage inside container:
     # GPU mode:
@@ -47,11 +47,6 @@ def make_face_image(size: int = 640) -> Image.Image:
     return img
 
 
-def make_random_image(width: int = 224, height: int = 224) -> Image.Image:
-    rng = np.random.default_rng(42)
-    return Image.fromarray(rng.integers(0, 256, (height, width, 3), dtype=np.uint8), "RGB")
-
-
 def _stats(times_s: list[float]) -> dict:
     arr = np.array(times_s) * 1000  # ms
     return {
@@ -119,61 +114,6 @@ def bench_insightface(n_warmup: int = 5, n_runs: int = 30) -> None:
     print(f"  320×320 median  : {s2['median_ms']:.1f} ms  ({s2['ips']:.1f} img/s)")
 
 
-def bench_siglip(
-    n_warmup: int = 3,
-    n_runs: int = 20,
-    batch_sizes: tuple = (1, 4, 8, 16, 32),
-) -> None:
-    import torch
-
-    import winnow.embeddings as emb_mod
-    emb_mod._siglip_model = None
-    emb_mod._siglip_processor = None
-    emb_mod._siglip_loaded = False
-
-    print("  Loading model...")
-    t_load = time.perf_counter()
-    model, processor = emb_mod.get_siglip_model()
-    load_s = time.perf_counter() - t_load
-
-    if model is None:
-        print("  SKIP: SigLIP failed to load")
-        return
-
-    device = next(model.parameters()).device
-    print(f"  Model load time : {load_s:.2f} s  (device: {device})")
-
-    print(f"  {'Batch':>5}  {'ms/batch':>10}  {'ms/img':>8}  {'img/s':>8}  {'p95/img':>9}")
-    for bs in batch_sizes:
-        imgs = [make_random_image(224, 224) for _ in range(bs)]
-        inputs = processor(images=imgs, return_tensors="pt")
-        inputs = {k: v.to(device) for k, v in inputs.items()}
-
-        # Warmup
-        for _ in range(n_warmup):
-            with torch.no_grad():
-                model(**inputs)
-        if str(device) != "cpu":
-            torch.cuda.synchronize()
-
-        times: list[float] = []
-        for _ in range(n_runs):
-            if str(device) != "cpu":
-                torch.cuda.synchronize()
-            t0 = time.perf_counter()
-            with torch.no_grad():
-                model(**inputs)
-            if str(device) != "cpu":
-                torch.cuda.synchronize()
-            times.append(time.perf_counter() - t0)
-
-        s = _stats(times)
-        print(
-            f"  {bs:>5}  {s['median_ms']:>10.1f}  {s['median_ms']/bs:>8.2f}"
-            f"  {bs * 1000 / s['median_ms']:>8.1f}  {s['p95_ms']/bs:>9.2f}"
-        )
-
-
 def main() -> None:
     print("=" * 56)
     print("  winnow inference benchmark")
@@ -185,10 +125,6 @@ def main() -> None:
     bench_insightface()
     print()
 
-    print("── SigLIP google/siglip-base-patch16-224  (objects) ───")
-    bench_siglip()
-    print()
-
 
 if __name__ == "__main__":
     # Add winnow to path when run directly inside container

tests/test_config.py

@@ -18,10 +18,10 @@ def test_config_loads_defaults(monkeypatch):
     assert cfg.OUTPUT_DIR == "./frigate_train"
     assert cfg.YEARS_FILTER == 10
     assert cfg.MIN_FACE_WIDTH == 90
-    assert cfg.MIN_FACE_COUNT == 0
+    assert cfg.MIN_FACE_COUNT == 3
     assert cfg.BLUR_THRESHOLD == 120.0
     assert cfg.MIN_CONFIDENCE == 0.7
-    assert cfg.MAX_AUTO_IMAGES == 80
+    assert cfg.MAX_AUTO_IMAGES == 20
     assert cfg.QUALITY_REPLACEMENT is True
     assert cfg.FACE_MARGIN == 0.15
     assert cfg.USE_FULL_RESOLUTION is True

winnow/__init__.py

@@ -1,8 +1,7 @@
 """Immich to Frigate training set curator.
 
 AI-powered tool to extract high-quality, diverse training images from your
-Immich library for Frigate's Face Recognition (ArcFace) and Object/State
+Immich library for Frigate's face recognition (ArcFace/Buffalo_L).
-Classification models.
 """
 
 from importlib.metadata import PackageNotFoundError, version

winnow/cache.py

@@ -15,7 +15,6 @@ logger = logging.getLogger(__name__)
 # Model versions — bump these when the upstream model changes
 MODEL_VERSIONS = {
     "insightface": "buffalo_l_v1",
-    "siglip": "siglip-base-patch16-224_v1",
     "immich": "immich_buffalo_l_v1",
 }
 

winnow/cli.py

@@ -9,7 +9,7 @@ from rich.prompt import Confirm
 
 from .config import Config, ConfigManager
 from .executor import execute_jobs, upload_to_frigate
-from .immich_api import get_people
+from .immich_api import get_people, merge_people
 from .jobs import _show_preview, auto_configure, interactive_configure
 from .log_config import console, setup_logging
 from .upload_tracker import find_by_crop_dimension, get_person_summary, reset_person
@@ -51,6 +51,97 @@ def _handle_trace_crop(size_str: str) -> None:
     sys.exit(0)
 
 
+def _handle_duplicate_people(people: list[dict]) -> list[dict]:
+    """Warn about or merge Immich people that share the same name.
+
+    Duplicates arise when Immich creates separate person records for the same
+    individual (e.g. unmerged face clusters). Without handling, winnow would
+    run multiple jobs for the same Frigate folder and overwrite its own output,
+    leaving far fewer training images than expected.
+
+    With MERGE_DUPLICATE_PEOPLE=false (default): prints a warning, skips the
+    smaller duplicates so only the person with the most assets is processed,
+    and returns a deduplicated people list.
+    With MERGE_DUPLICATE_PEOPLE=true: merges each duplicate group inside
+    Immich via its API (permanently combines the face records), then
+    re-fetches the people list so the rest of the run sees the merged state.
+    """
+    from collections import defaultdict
+
+    by_name: dict[str, list[dict]] = defaultdict(list)
+    for p in people:
+        name = (p.get("name") or "").strip()
+        if name:
+            by_name[name].append(p)
+
+    duplicates = {name: ps for name, ps in by_name.items() if len(ps) > 1}
+    if not duplicates:
+        return people
+
+    if not Config.MERGE_DUPLICATE_PEOPLE:
+        rprint("\n[bold yellow]⚠  Duplicate person names detected in Immich:[/bold yellow]")
+        for name, ps in sorted(duplicates.items()):
+            ordered = sorted(ps, key=lambda x: x.get("assetCount", 0), reverse=True)
+            entries = ", ".join(
+                f"[dim]{p['id'][:8]}…[/dim] ({p.get('assetCount', 0)} assets)"
+                for p in ordered
+            )
+            rprint(f"  [yellow]{name}[/yellow] → {len(ps)} people: {entries}")
+            skipped = ordered[1:]
+            rprint(
+                f"  [dim]  Processing largest only "
+                f"({ordered[0].get('assetCount', 0)} assets). "
+                f"Skipping {len(skipped)} smaller duplicate(s) to avoid overwriting output.[/dim]"
+            )
+        rprint(
+            "  [dim]Set MERGE_DUPLICATE_PEOPLE=true to permanently merge duplicates "
+            "inside Immich (keeps the person with the most assets).[/dim]\n"
+        )
+        # Return deduplicated list — keep only the largest per name so that
+        # downstream job creation never runs two jobs for the same Frigate folder.
+        skip_ids = {
+            p["id"]
+            for ps in duplicates.values()
+            for p in sorted(ps, key=lambda x: x.get("assetCount", 0), reverse=True)[1:]
+        }
+        return [p for p in people if p["id"] not in skip_ids]
+
+    # Auto-merge: survivor = largest asset count, rest merge into it inside Immich
+    merged_any = False
+    for name, ps in sorted(duplicates.items()):
+        ordered = sorted(ps, key=lambda x: x.get("assetCount", 0), reverse=True)
+        survivor = ordered[0]
+        merge_ids = [p["id"] for p in ordered[1:]]
+        rprint(
+            f"  [cyan]Merging {name!r} inside Immich:[/cyan] keeping "
+            f"[dim]{survivor['id'][:8]}…[/dim] ({survivor.get('assetCount', 0)} assets), "
+            f"absorbing {len(merge_ids)} smaller duplicate(s)..."
+        )
+        if merge_people(survivor["id"], merge_ids):
+            rprint(f"  [green]✓ Merged {name!r}[/green]")
+            merged_any = True
+        else:
+            rprint(f"  [red]✗ Failed to merge {name!r}[/red]")
+
+    if merged_any:
+        rprint("  [dim]Re-fetching people after merge...[/dim]")
+        return get_people()
+
+    return people
+
+
+_UNSUPPORTED_VARS = [
+    "ENABLE_FRIGATE_SCORES",
+    "FRIGATE_SCORE_CEILING",
+    "MIN_FACE_WIDTH",
+    "FACE_MARGIN",
+    "ENABLE_FACE_ALIGNMENT",
+    "USE_FULL_RESOLUTION",
+    "MIN_CONFIDENCE",
+    "BLUR_THRESHOLD",
+]
+
+
 def main() -> None:
     """Entry point for winnow CLI."""
     try:
@@ -66,6 +157,17 @@ def main() -> None:
     [dim]Immich -> Frigate Training Data Curator[/dim]
         """)
 
+        set_unsupported = [v for v in _UNSUPPORTED_VARS if os.environ.get(v)]
+        if set_unsupported:
+            console.print(
+                f"[bold yellow]⚠  Advanced tuning vars set: "
+                f"{', '.join(set_unsupported)}[/bold yellow]"
+            )
+            console.print(
+                "[dim]  These defaults are calibrated for Frigate's ArcFace requirements. "
+                "Image quality issues caused by non-default values will not be investigated.[/dim]\n"
+            )
+
         ConfigManager.get().interactive_setup()
 
         try:
@@ -77,9 +179,25 @@ def main() -> None:
         rprint(f"Server: [dim]{Config.IMMICH_URL}[/dim]")
         rprint(f"Output: [dim]{Config.OUTPUT_DIR}[/dim]")
 
-        # Handle RESET_PERSON before anything else
+        # Handle RESET_PERSON before anything else.
+        # RESET_PERSON=* resets every tracked person; any other value resets
+        # that specific person by name.
         reset_person_name = os.environ.get("RESET_PERSON", "").strip()
         if reset_person_name:
+            if reset_person_name == "*":
+                names = list(get_person_summary().keys())
+                if "*" in names:
+                    rprint(
+                        "[yellow]Note: a person literally named '*' exists in the tracker "
+                        "and will be reset along with everyone else.[/yellow]"
+                    )
+                if names:
+                    for name in names:
+                        reset_person(name)
+                    rprint(f"[bold yellow]Reset tracking data for all {len(names)} people.[/bold yellow]")
+                else:
+                    rprint("[dim]No tracking data to reset.[/dim]")
+            else:
                 reset_person(reset_person_name)
                 rprint(f"[bold yellow]Reset tracking data for: {reset_person_name}[/bold yellow]")
 
@@ -103,6 +221,8 @@ def main() -> None:
             rprint("[bold red]Could not fetch people from Immich. Check URL/Key.[/bold red]")
             return
 
+        people = _handle_duplicate_people(people)
+
         # Auto mode when no TTY (Docker, cron, pipes) — the primary use case.
         # A TTY means local interactive use; AUTO_MODE=true overrides that for scripting.
         auto_mode = not sys.stdin.isatty() or os.environ.get("AUTO_MODE", "").lower() in ("true", "1", "yes")

winnow/config.py

@@ -29,13 +29,14 @@ class _Config:
     MIN_FACE_WIDTH: int = 90
     BLUR_THRESHOLD: float = 120.0
     MIN_CONFIDENCE: float = 0.7
-    MAX_AUTO_IMAGES: int = 80
+    MAX_AUTO_IMAGES: int = 20
     QUALITY_REPLACEMENT: bool = True
-    FRIGATE_SCORE_CEILING: float = 0.0
+    FRIGATE_SCORE_CEILING: float | None = None
     ENABLE_FRIGATE_SCORES: bool = True
 
     # People filtering
-    MIN_FACE_COUNT: int = 0
+    MIN_FACE_COUNT: int = 3
+    MERGE_DUPLICATE_PEOPLE: bool = False
 
     # Output quality
     FACE_MARGIN: float = 0.15
@@ -59,12 +60,14 @@ class _Config:
         self.OUTPUT_DIR = os.getenv("OUTPUT_DIR", "./frigate_train")
         self.YEARS_FILTER = int(os.getenv("YEARS_FILTER", "10"))
         self.MIN_FACE_WIDTH = int(os.getenv("MIN_FACE_WIDTH", "90"))
-        self.MIN_FACE_COUNT = int(os.getenv("MIN_FACE_COUNT", "0"))
+        self.MIN_FACE_COUNT = int(os.getenv("MIN_FACE_COUNT", "3"))
+        self.MERGE_DUPLICATE_PEOPLE = os.getenv("MERGE_DUPLICATE_PEOPLE", "false").lower() in ("true", "1", "yes")
         self.BLUR_THRESHOLD = float(os.getenv("BLUR_THRESHOLD", "120.0"))
         self.MIN_CONFIDENCE = float(os.getenv("MIN_CONFIDENCE", "0.7"))
-        self.MAX_AUTO_IMAGES = int(os.getenv("MAX_AUTO_IMAGES", "80"))
+        self.MAX_AUTO_IMAGES = int(os.getenv("MAX_AUTO_IMAGES", "20"))
         self.QUALITY_REPLACEMENT = os.getenv("QUALITY_REPLACEMENT", "true").lower() in ("true", "1", "yes")
-        self.FRIGATE_SCORE_CEILING = float(os.getenv("FRIGATE_SCORE_CEILING", "0.0"))
+        _ceiling_env = os.getenv("FRIGATE_SCORE_CEILING", "").strip()
+        self.FRIGATE_SCORE_CEILING = float(_ceiling_env) if _ceiling_env else None
         self.ENABLE_FRIGATE_SCORES = os.getenv("ENABLE_FRIGATE_SCORES", "true").lower() in ("true", "1", "yes")
         self.FACE_MARGIN = float(os.getenv("FACE_MARGIN", "0.15"))
         self.USE_FULL_RESOLUTION = os.getenv("USE_FULL_RESOLUTION", "true").lower() in ("true", "1", "yes")

winnow/diversity.py

@@ -5,7 +5,7 @@ Selection pipeline:
 1. Concurrent thumbnail download
 2. Quality filtering (blur, IR, exposure, confidence, face size)
 3. Face crop extraction (embed person's face, not full image)
-4. Embedding computation (InsightFace or SigLIP)
+4. Embedding computation (InsightFace)
 5. Cluster-aware selection (K-Medoids + FPS with hard example weighting)
 """
 
@@ -28,7 +28,6 @@ def select_diverse_assets(
     limit: int | str,
     entity_name: str,
     selection_mode: str = "smart",
-    entity_type: str = "face",
     person_id: str | None = None,
     progress_callback=None,
 ) -> list:
@@ -38,9 +37,8 @@ def select_diverse_assets(
     Args:
         assets: List of asset dicts from Immich API
         limit: Number to select, or "auto" for dynamic selection
-        entity_name: Name of the person/object for logging
+        entity_name: Name of the person for logging
         selection_mode: 'smart' (embedding-based) or 'time' (time spread)
-        entity_type: 'face' or 'object' - determines embedding model
         progress_callback: Optional callback(current, total) for progress
 
     Returns:
@@ -53,14 +51,13 @@ def select_diverse_assets(
     # Sort by creation time
     assets = sorted(assets, key=lambda x: x.get("fileCreatedAt", ""))
 
-    if selection_mode != "smart" or not is_embedding_available(entity_type):
+    if selection_mode != "smart" or not is_embedding_available():
         if selection_mode == "smart":
-            model_name = "InsightFace" if entity_type == "face" else "SigLIP"
+            logger.warning("InsightFace unavailable. Falling back to time spread.")
-            logger.warning(f"{model_name} unavailable. Falling back to time spread.")
         return _select_time_spread(assets, limit)
 
     try:
-        return _select_by_embedding(assets, limit, entity_type, person_id, progress_callback)
+        return _select_by_embedding(assets, limit, person_id, progress_callback)
     except Exception as e:
         logger.error(f"Smart Diversity failed: {e}. Falling back to time spread.")
         return _select_time_spread(assets, limit)
@@ -179,7 +176,6 @@ def _crop_face_from_thumbnail(
 def _select_by_embedding(
     assets: list,
     limit: int | str,
-    entity_type: str,
     person_id: str | None = None,
     progress_callback=None,
 ) -> list:
@@ -188,7 +184,7 @@ def _select_by_embedding(
     Pipeline:
     1. Concurrent thumbnail download
     2. Quality filtering
-    3. Face crop extraction (face mode only)
+    3. Face crop extraction
     4. Embedding computation
     5. Cluster-aware selection with hard example weighting
     """
@@ -243,7 +239,6 @@ def _select_by_embedding(
 
             confidence = _get_face_confidence(asset, person_id=person_id)
 
-            if entity_type == "face":
             face_bbox = _get_face_bbox(asset, person_id=person_id)
             quality = assess_quality(
                 img,
@@ -261,10 +256,8 @@ def _select_by_embedding(
             asset["quality_score"] = quality.blur_score
             face_crop = _crop_face_from_thumbnail(img, asset, person_id=person_id)
             embed_img = face_crop if face_crop is not None else img
-            else:
-                embed_img = img
 
-            emb = get_embedding(embed_img, entity_type, asset_id=asset["id"])
+            emb = get_embedding(embed_img, asset_id=asset["id"])
             if emb is not None:
                 embeddings.append(emb)
                 valid_candidates.append(asset)
@@ -281,16 +274,86 @@ def _select_by_embedding(
         logger.warning(f"Only {len(valid_candidates)} valid embeddings. Returning all.")
         return valid_candidates
 
-    # --- Phase 5: Cluster-aware selection ---
+    # --- Phase 5: Near-duplicate removal ---
+    # Burst shots and repeated near-identical photos produce embeddings that are
+    # close but not identical, so FPS doesn't filter them out on its own.
+    # Greedily drop any candidate within DEDUP_THRESHOLD cosine distance of a
+    # higher-quality image already in the kept set.
+    embeddings, valid_candidates, confidence_scores = _dedup_embeddings(
+        embeddings, valid_candidates, confidence_scores
+    )
+
+    # Re-check after dedup: pool may have shrunk below limit
+    if limit != "auto" and len(valid_candidates) < limit:
+        logger.warning(f"Only {len(valid_candidates)} embeddings after near-duplicate removal. Returning all.")
+        return valid_candidates
+
+    # --- Phase 6: Cluster-aware selection ---
     return _cluster_aware_selection(
         embeddings,
         valid_candidates,
         limit,
-        entity_type=entity_type,
         confidence_scores=confidence_scores,
     )
 
 
+# =============================================================================
+# Near-Duplicate Removal
+# =============================================================================
+
+_DEDUP_THRESHOLD = 0.20  # cosine distance — burst shots ~0.01-0.05, same-event similar shots ~0.10-0.20
+
+
+def _dedup_embeddings(
+    embeddings: list,
+    candidates: list,
+    confidence_scores: list,
+) -> tuple[list, list, list]:
+    """Greedy near-duplicate removal before clustering.
+
+    Sorts by quality score descending (best first), then for each candidate
+    drops it if any already-kept embedding is within _DEDUP_THRESHOLD cosine
+    distance. This eliminates burst-shot near-duplicates while preserving the
+    highest-quality representative from each near-identical group.
+    """
+    if len(embeddings) < 2:
+        return embeddings, candidates, confidence_scores
+
+    emb_matrix = np.vstack(embeddings)
+    norms = np.linalg.norm(emb_matrix, axis=1, keepdims=True)
+    emb_normed = emb_matrix / np.maximum(norms, 1e-8)
+
+    # Sort by quality descending so the best image in each near-duplicate group wins.
+    # Use explicit None check so a legitimate quality_score=0.0 isn't treated as missing.
+    quality_scores = [qs if (qs := c.get("quality_score")) is not None else 0.0 for c in candidates]
+    order = sorted(range(len(candidates)), key=lambda i: quality_scores[i], reverse=True)
+
+    kept_indices = []
+    # Pre-allocate a max-size buffer and fill row-by-row — eliminates the O(K²)
+    # copy overhead from vstack-on-keep while keeping identical arithmetic.
+    kept_buf = np.empty((len(order), emb_normed.shape[1]), dtype=emb_normed.dtype)
+    n_kept = 0
+
+    for i in order:
+        if n_kept > 0:
+            sims = emb_normed[i] @ kept_buf[:n_kept].T
+            if np.any(sims > 1 - _DEDUP_THRESHOLD):
+                continue
+        kept_buf[n_kept] = emb_normed[i]
+        n_kept += 1
+        kept_indices.append(i)
+
+    dropped = len(embeddings) - len(kept_indices)
+    if dropped:
+        logger.info(f"Near-duplicate removal dropped {dropped} images (threshold {_DEDUP_THRESHOLD}).")
+
+    return (
+        [embeddings[i] for i in kept_indices],
+        [candidates[i] for i in kept_indices],
+        [confidence_scores[i] for i in kept_indices],
+    )
+
+
 # =============================================================================
 # K-Medoids (Lightweight Implementation)
 # =============================================================================
@@ -322,7 +385,7 @@ def _kmedoids(dist_matrix: np.ndarray, k: int, max_iter: int = 50) -> tuple[list
     # Iterative swap step
     medoids = list(medoids)
     labels = np.argmin(dist_matrix[:, medoids], axis=1)
-    cost = sum(dist_matrix[i, medoids[labels[i]]] for i in range(n))
+    cost = dist_matrix[np.arange(n), np.array(medoids)[labels]].sum()
 
     for _ in range(max_iter):
         improved = False
@@ -337,7 +400,7 @@ def _kmedoids(dist_matrix: np.ndarray, k: int, max_iter: int = 50) -> tuple[list
                 new_medoids = medoids.copy()
                 new_medoids[m_idx] = cand
                 new_labels = np.argmin(dist_matrix[:, new_medoids], axis=1)
-                new_cost = sum(dist_matrix[i, new_medoids[new_labels[i]]] for i in range(n))
+                new_cost = dist_matrix[np.arange(n), np.array(new_medoids)[new_labels]].sum()
                 if new_cost < cost:
                     medoids = new_medoids
                     labels = new_labels
@@ -358,11 +421,11 @@ def _kmedoids(dist_matrix: np.ndarray, k: int, max_iter: int = 50) -> tuple[list
 # =============================================================================
 
 
-def _compute_adaptive_threshold(emb_normed: np.ndarray, entity_type: str) -> float:
+def _compute_adaptive_threshold(emb_normed: np.ndarray) -> float:
     """Compute adaptive FPS stop threshold based on actual embedding distribution.
 
     Instead of a hardcoded threshold, samples pairwise distances and sets
-    the threshold as a fraction of the median pairwise distance.
+    the threshold as 20% of the median pairwise distance.
     """
     n = len(emb_normed)
     sample_size = min(200, n)
@@ -370,20 +433,14 @@ def _compute_adaptive_threshold(emb_normed: np.ndarray, entity_type: str) -> flo
     indices = rng.choice(n, sample_size, replace=False) if n > sample_size else np.arange(n)
     sample = emb_normed[indices]
 
-    # Compute pairwise cosine distances for the sample
     pairwise = 1 - sample @ sample.T
     upper_tri = pairwise[np.triu_indices(len(sample), k=1)]
+    if len(upper_tri) == 0:
+        return 0.05
     median_dist = float(np.median(upper_tri))
+    threshold = max(0.05, median_dist * 0.20)
 
-    # Faces: 20% of median (tighter — want fewer, more distinct images)
+    logger.debug(f"Adaptive threshold: {threshold:.4f} (median_dist={median_dist:.4f})")
-    # Objects: 10% of median (wider — want more diversity)
-    fraction = 0.20 if entity_type == "face" else 0.10
-    threshold = max(0.05, median_dist * fraction)
-
-    logger.debug(
-        f"Adaptive threshold: {threshold:.4f} "
-        f"(median_dist={median_dist:.4f}, fraction={fraction}, type={entity_type})"
-    )
     return threshold
 
 
@@ -391,7 +448,6 @@ def _cluster_aware_selection(
     embeddings: list,
     candidates: list,
     limit: int | str,
-    entity_type: str = "face",
     confidence_scores: list | None = None,
 ) -> list:
     """Two-stage selection: K-Medoids clustering → FPS with hard example weighting.
@@ -411,17 +467,17 @@ def _cluster_aware_selection(
 
     # Build confidence weight array for hard example boosting
     conf_array = np.ones(n)
-    if confidence_scores and entity_type == "face":
+    if confidence_scores:
         for i, c in enumerate(confidence_scores):
             if c is not None:
                 conf_array[i] = c
 
     # Compute adaptive threshold for auto mode
-    auto_threshold = _compute_adaptive_threshold(emb_normed, entity_type) if limit == "auto" else 0.0
+    auto_threshold = _compute_adaptive_threshold(emb_normed) if limit == "auto" else 0.0
     target = Config.MAX_AUTO_IMAGES if limit == "auto" else limit
 
     # --- Stage 1: K-Medoids clustering ---
-    k = min(max(5, target // 4), n // 3, n)  # e.g., 5-20 clusters
+    k = min(max(5, target // 4), max(1, n // 3), n)  # e.g., 1-20 clusters
     logger.debug(f"Clustering {n} embeddings into {k} groups (K-Medoids)...")
 
     # Compute full cosine distance matrix
@@ -469,15 +525,9 @@ def _cluster_aware_selection(
         min_dists = np.minimum(min_dists, dists_to_new)
         min_dists[best_idx] = -np.inf
 
-    # Log hard example stats
-    if entity_type == "face":
     selected_conf = [conf_array[i] for i in selected if conf_array[i] < 1.0]
     hard_count = sum(1 for c in selected_conf if c < 0.85)
-        logger.info(
+    logger.info(f"Selection complete: {len(selected)} images ({hard_count} hard examples with confidence < 0.85).")
-            f"Selection complete: {len(selected)} images " f"({hard_count} hard examples with confidence < 0.85)."
-        )
-    else:
-        logger.info(f"Selection complete: {len(selected)} diverse images.")
 
     return [candidates[i] for i in selected]
 

winnow/embeddings.py

@@ -1,8 +1,7 @@
 """
-Unified embedding interface for faces and objects.
+Embedding interface for face diversity selection.
 
 - Faces: InsightFace (ArcFace/Buffalo_L) — or reuse from Immich
-- Objects: SigLIP (Vision Transformer via transformers)
 - Caching: Disk-based cache avoids recomputation on reruns
 """
 
@@ -42,12 +41,9 @@ def _suppress_output():
         os.close(saved_err)
 
 
-# Lazy-loaded singletons
+# Lazy-loaded singleton
 _insightface_app = None
 _insightface_loaded = False
-_siglip_model = None
-_siglip_processor = None
-_siglip_loaded = False
 
 
 def _is_force_cpu() -> bool:
@@ -202,169 +198,33 @@ def get_face_embedding(img_pil: Image.Image) -> np.ndarray | None:
 
 
 # =============================================================================
-# SigLIP (Objects)
+# Embedding Interface with Caching
-# =============================================================================
-
-
-def get_siglip_model():
-    """Singleton for SigLIP model and processor with GPU auto-detection."""
-    global _siglip_model, _siglip_processor, _siglip_loaded
-    if _siglip_loaded:
-        return _siglip_model, _siglip_processor
-    _siglip_loaded = True
-
-    try:
-        import warnings
-
-        import torch
-        from transformers import AutoImageProcessor, SiglipVisionModel
-
-        model_name = "google/siglip-base-patch16-224"
-
-        # Disk cache check — path derived from model_name using HuggingFace's slug convention
-        hf_home = os.environ.get("HF_HOME", os.path.join(os.path.expanduser("~"), ".cache", "huggingface"))
-        cache_slug = "models--" + model_name.replace("/", "--")
-        model_cache = Path(hf_home) / "hub" / cache_slug
-        if model_cache.exists() and any(model_cache.iterdir()):
-            logger.info(f"SigLIP {model_name}: found in model cache")
-        else:
-            logger.info(f"SigLIP {model_name}: not cached — downloading now (~380 MB)")
-
-        logger.info(f"SigLIP {model_name}: loading into memory...")
-        t0 = time.time()
-
-        with warnings.catch_warnings():
-            warnings.filterwarnings("ignore", category=FutureWarning)
-            warnings.filterwarnings("ignore", message=".*use_fast.*")
-            _siglip_processor = AutoImageProcessor.from_pretrained(model_name, use_fast=True)
-            _siglip_model = SiglipVisionModel.from_pretrained(model_name)
-
-        _siglip_model.eval()
-
-        # Move to GPU if available (ROCm builds expose torch.cuda.is_available() == True)
-        if not _is_force_cpu():
-            if torch.cuda.is_available():
-                _siglip_model = _siglip_model.cuda()
-                device_name = "CUDA GPU"
-            elif hasattr(torch, "xpu") and torch.xpu.is_available():
-                _siglip_model = _siglip_model.to("xpu")
-                device_name = "Intel XPU"
-            elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
-                _siglip_model = _siglip_model.to("mps")
-                device_name = "Apple MPS"
-            else:
-                device_name = "CPU"
-        else:
-            device_name = "CPU (FORCE_CPU)"
-
-        logger.info(f"SigLIP {model_name}: ready on {device_name} ({time.time() - t0:.1f}s)")
-        return _siglip_model, _siglip_processor
-
-    except ImportError as e:
-        logger.error(f"transformers/torch not installed: {e}")
-        return None, None
-    except Exception as e:
-        logger.error(f"Failed to load SigLIP: {e}")
-        return None, None
-
-
-def get_object_embedding(img_pil: Image.Image) -> np.ndarray | None:
-    """Get 768-dim SigLIP embedding for an image."""
-    model, processor = get_siglip_model()
-    if model is None:
-        return None
-
-    try:
-        import torch
-
-        inputs = processor(images=img_pil, return_tensors="pt")
-        device = next(model.parameters()).device
-        inputs = {k: v.to(device) for k, v in inputs.items()}
-
-        with torch.no_grad():
-            outputs = model(**inputs)
-            return outputs.pooler_output.squeeze().cpu().numpy()
-    except Exception as e:
-        logger.error(f"Error getting object embedding: {e}")
-        return None
-
-
-def get_object_embeddings_batch(images: list[Image.Image]) -> list[np.ndarray | None]:
-    """Get SigLIP embeddings for a batch of images (GPU-efficient)."""
-    model, processor = get_siglip_model()
-    if model is None:
-        return [None] * len(images)
-
-    try:
-        import torch
-
-        inputs = processor(images=images, return_tensors="pt", padding=True)
-        device = next(model.parameters()).device
-        inputs = {k: v.to(device) for k, v in inputs.items()}
-
-        with torch.no_grad():
-            outputs = model(**inputs)
-            embeddings = outputs.pooler_output.cpu().numpy()
-            return [embeddings[i] for i in range(len(embeddings))]
-    except Exception as e:
-        logger.error(f"Error in batch embedding: {e}")
-        # Fall back to individual computation
-        return [get_object_embedding(img) for img in images]
-
-
-# =============================================================================
-# Unified Interface with Caching
 # =============================================================================
 
 
 def get_embedding(
     img_pil: Image.Image,
-    entity_type: str = "face",
     asset_id: str | None = None,
-    immich_embedding: np.ndarray | None = None,
 ) -> np.ndarray | None:
-    """Get embedding for an image based on entity type.
+    """Get embedding for a face image.
 
-    Priority:
+    Checks disk cache first (if enabled and asset_id provided),
-    1. Pre-fetched Immich embedding (if provided)
+    then falls back to local InsightFace computation.
-    2. Disk cache (if enabled and asset_id provided)
-    3. Local model computation (InsightFace or SigLIP)
-
-    Args:
-        img_pil: The image to embed
-        entity_type: 'face' or 'object'
-        asset_id: Optional asset ID for cache lookup
-        immich_embedding: Optional pre-fetched embedding from Immich API
     """
     from .config import Config
 
     use_cache = Config.ENABLE_CACHE and asset_id is not None
     cache = get_cache(Config.CACHE_DIR) if use_cache else None
-    # Use a single consistent cache key per model so lookups and stores always match.
-    # "immich" was previously used as the face key on the lookup path but "insightface"
-    # on the store path — meaning the cache was never hit for locally-computed embeddings.
-    cache_key = "insightface" if entity_type == "face" else "siglip"
 
-    # 1. Use Immich embedding if provided
-    if immich_embedding is not None:
     if cache:
-            cache.put(asset_id, immich_embedding, cache_key)
+        cached = cache.get(asset_id, "insightface")
-        return immich_embedding
-
-    # 2. Check disk cache
-    if cache:
-        cached = cache.get(asset_id, cache_key)
         if cached is not None:
             return cached
 
-    # 3. Compute locally
-    if entity_type == "face":
     emb = get_face_embedding(img_pil)
-    else:
-        emb = get_object_embedding(img_pil)
 
     if emb is not None and cache:
-        cache.put(asset_id, emb, cache_key)
+        cache.put(asset_id, emb, "insightface")
 
     return emb
 
@@ -378,36 +238,18 @@ def _is_module_available(module_name: str) -> bool:
         return False
 
 
-def is_embedding_available(entity_type: str = "face", *, load: bool = False) -> bool:
+def is_embedding_available(*, load: bool = False) -> bool:
-    """Check if embedding model is available for the given entity type.
+    """Check if InsightFace is available.
 
     By default this performs a lightweight import-check only (no model loading).
-    Pass ``load=True`` to actually load the model (expensive, hundreds of MB).
+    Pass ``load=True`` to actually load the model (expensive, ~300 MB).
-
-    Args:
-        entity_type: 'face' or 'object'
-        load: If True, fully load the model to verify. If False (default),
-              only check that the required packages are importable.
     """
     if load:
-        if entity_type == "face":
         return get_insightface_app() is not None
-        model, _ = get_siglip_model()
-        return model is not None
-
-    # Lightweight check: just verify the packages are importable
-    if entity_type == "face":
     return _is_module_available("insightface") and _is_module_available("onnxruntime")
-    return _is_module_available("transformers") and _is_module_available("torch")
 
 
-def load_embedding_model(entity_type: str = "face") -> bool:
+def load_embedding_model() -> bool:
-    """Explicitly load the embedding model for the given entity type.
+    """Explicitly load InsightFace. Returns True if the model loaded successfully."""
-
-    Returns True if the model loaded successfully.
-    """
-    if entity_type == "face":
     return get_insightface_app() is not None
-    model, _ = get_siglip_model()
-    return model is not None
 

winnow/executor.py

@@ -19,7 +19,7 @@ from .frigate_api import (
     get_frigate_person_files,
     recognize_face,
 )
-from .image_processing import process_face_mode, process_full_mode, process_object_mode
+from .image_processing import process_face_mode
 from .immich_api import fetch_face_data, fetch_full_image
 from .log_config import console
 from .quality import assess_quality
@@ -31,13 +31,29 @@ from .upload_tracker import (
     has_frigate_scores,
     mark_rejected,
     mark_uploaded,
-    record_frigate_file,
+    record_frigate_files_batch,
     remove_frigate_file,
 )
 
 logger = logging.getLogger(__name__)
 
 
+def _safe_person_dir(output_dir: str, person_name: str) -> str:
+    """Return the output subdirectory for a person, raising ValueError on path traversal.
+
+    os.path.join silently discards output_dir when person_name is absolute,
+    and '../..' sequences resolve outside the tree. Both are rejected here.
+    """
+    candidate = os.path.realpath(os.path.join(output_dir, person_name))
+    base = os.path.realpath(output_dir)
+    # Use the base path as its own prefix when it's the filesystem root ("/"),
+    # otherwise append os.sep — avoids the false "//" double-slash when base == "/".
+    base_prefix = base if base == os.sep else base + os.sep
+    if not candidate.startswith(base_prefix) and candidate != base:
+        raise ValueError(f"Person name {person_name!r} escapes output directory — skipping")
+    return candidate
+
+
 def _reconcile_frigate_mappings(
     person_name: str,
     known_files_before: set[str],
@@ -84,10 +100,12 @@ def _reconcile_frigate_mappings(
             except (ValueError, IndexError):
                 return 0.0
 
-        for (fname, asset_id), frigate_file in zip(uploaded, sorted(new_files, key=_ts)):
+        mappings = {
-            if asset_id:
+            frigate_file: asset_id
-                record_frigate_file(person_name, frigate_file, asset_id)
+            for (_, asset_id), frigate_file in zip(uploaded, sorted(new_files, key=_ts))
-        logger.debug(f"{person_name}: batch-mapped {target} Frigate file(s)")
+            if asset_id
+        }
+        record_frigate_files_batch(person_name, mappings)
     elif len(new_files) > target:
         logger.info(
             f"{person_name}: {len(new_files)} new Frigate files for {target} uploads"
@@ -155,6 +173,18 @@ def execute_jobs(jobs: list[dict]) -> None:
 
     use_full_res = Config.USE_FULL_RESOLUTION
 
+    # Load InsightFace app for landmark-based crop alignment.
+    # The model is already resident from the diversity/embedding phase, so this
+    # is just a singleton lookup — no load cost.
+    insightface_app = None
+    if Config.ENABLE_FACE_ALIGNMENT:
+        try:
+            from .embeddings import get_insightface_app
+
+            insightface_app = get_insightface_app()
+        except Exception as e:
+            logger.debug(f"InsightFace unavailable for crop alignment: {e}")
+
     with Progress(
         SpinnerColumn(),
         TextColumn("[progress.description]{task.description}"),
@@ -166,14 +196,17 @@ def execute_jobs(jobs: list[dict]) -> None:
         overall_task = progress.add_task("[green]Overall Progress", total=grand_total)
 
         for job in jobs:
-            person, assets, config = job["person"], job["assets"], job["config"]
+            person, assets = job["person"], job["assets"]
-            name, mode = person["name"], config.get("mode", "face")
+            name = person["name"]
 
             job_task = progress.add_task(f"Processing {name}...", total=len(assets))
-            person_dir = os.path.join(Config.OUTPUT_DIR, name)
+            try:
+                person_dir = _safe_person_dir(Config.OUTPUT_DIR, name)
+            except ValueError as e:
+                logger.error(str(e))
+                continue
             # Face crops are transient (uploaded then discarded); wipe before each run.
-            # Object crops are the deliverable; preserve them across runs.
+            if os.path.isdir(person_dir):
-            if mode == "face" and os.path.isdir(person_dir):
                 shutil.rmtree(person_dir)
             os.makedirs(person_dir, exist_ok=True)
 
@@ -185,9 +218,8 @@ def execute_jobs(jobs: list[dict]) -> None:
             count = 0
             for asset in assets:
                 try:
-                    # For face mode, enrich the asset with face bounding box data
+                    # Enrich the asset with face bounding box data from the Immich
-                    # from the Immich faces API (not included in search/metadata results)
+                    # faces API (not included in search/metadata results).
-                    if mode == "face":
                     asset = _enrich_asset_with_face_data(asset, person)
                     # Skip download if detection confidence already disqualifies
                     # the asset — avoids fetching a large image we'll discard.
@@ -197,6 +229,7 @@ def execute_jobs(jobs: list[dict]) -> None:
                             f"[yellow]Skipped {asset['id']}"
                             f" (detection confidence {conf:.2f} < {Config.MIN_CONFIDENCE})[/yellow]"
                         )
+                        mark_rejected(asset["id"], person_name=name)
                         progress.advance(job_task)
                         progress.advance(overall_task)
                         continue
@@ -215,26 +248,21 @@ def execute_jobs(jobs: list[dict]) -> None:
                     if img is None:
                         progress.console.print(f"[red]Failed download {asset['id']}[/red]")
                     else:
-                        saved = (
+                        saved = process_face_mode(
-                            process_face_mode(img, asset, person, person_dir, count)
+                            img, asset, person, person_dir, count, insightface_app=insightface_app
-                            if mode == "face"
-                            else process_object_mode(img, config, person_dir, count)
-                            if mode == "object"
-                            else process_full_mode(img, person_dir, count)
                         )
                         if saved:
-                            # Record which asset produced which output file
                             filename = f"{count}.jpg"
                             asset_map[filename] = asset["id"]
                             score_map[filename] = asset.get("quality_score")
-                            if mode == "face" and isinstance(saved, tuple):
+                            if isinstance(saved, tuple):
                                 dims_map[filename] = saved
                             # Time-spread path: compute blur score from the downloaded
                             # image. Cap at 1440px so the scale matches the preview
                             # thumbnails the embedding path uses for scoring — Laplacian
                             # variance grows with resolution, making full-res and
                             # thumbnail scores incomparable if left uncapped.
-                            if mode == "face" and score_map[filename] is None:
+                            if score_map[filename] is None:
                                 try:
                                     score_img = img.convert("RGB") if img.mode != "RGB" else img
                                     if score_img.width > 1440 or score_img.height > 1440:
@@ -244,12 +272,6 @@ def execute_jobs(jobs: list[dict]) -> None:
                                 except Exception as exc:
                                     logger.debug(f"Quality score fallback for {asset['id']}: {exc}")
                                     score_map[filename] = 0.0  # unknown quality — treat as lowest
-                            # Also record object-mode variant filenames
-                            if mode == "object":
-                                for f in sorted(os.listdir(person_dir)):
-                                    if f.startswith(f"{count}_") and f not in asset_map:
-                                        asset_map[f] = asset["id"]
-                                        score_map[f] = asset.get("face_confidence")
 
                             count += 1
                         else:
@@ -277,25 +299,13 @@ def execute_jobs(jobs: list[dict]) -> None:
 def upload_to_frigate(jobs: list[dict]) -> None:
     """Upload processed face crops to Frigate via API with detailed logging.
 
-    Only runs for face-mode jobs. Object-mode crops are saved to the output
-    directory as the deliverable and must be copied to Frigate manually.
-
     After each successful upload, records the Immich asset ID in the
     upload tracker so it is skipped on future runs.
     """
-    face_jobs = [j for j in jobs if j["config"].get("mode", "face") == "face"]
+    if not jobs:
-
+        rprint("[dim]No jobs to upload.[/dim]")
-    if not face_jobs:
-        rprint("[dim]No face-mode jobs to upload.[/dim]")
         return
 
-    # Notify user about object-mode jobs that were skipped
-    object_jobs = [j for j in jobs if j["config"].get("mode") == "object"]
-    for job in object_jobs:
-        name = job["person"]["name"]
-        person_dir = os.path.join(Config.OUTPUT_DIR, name)
-        rprint(f"  [dim]📁 {name} (object): crops saved to {person_dir} — copy to Frigate manually[/dim]")
-
     frigate_url = os.environ.get("FRIGATE_URL", "")
     if not frigate_url:
         rprint("[yellow]⚠️  FRIGATE_URL not set, skipping upload.[/yellow]")
@@ -308,7 +318,7 @@ def upload_to_frigate(jobs: list[dict]) -> None:
     # from the asset_map stored on each job during execute_jobs()
     filename_to_asset_id: dict[str, dict[str, str]] = {}
     total_files = 0
-    for job in face_jobs:
+    for job in jobs:
         name = job["person"]["name"]
         asset_map = job.get("asset_map", {})
         filename_to_asset_id[name] = asset_map
@@ -318,7 +328,7 @@ def upload_to_frigate(jobs: list[dict]) -> None:
         rprint("  [yellow]No images found to upload.[/yellow]")
         return
 
-    rprint(f"  People: [bold]{len(face_jobs)}[/bold], Total images: [bold]{total_files}[/bold]")
+    rprint(f"  People: [bold]{len(jobs)}[/bold], Total images: [bold]{total_files}[/bold]")
 
     uploaded, failed = 0, 0
     max_retries = 2
@@ -336,14 +346,18 @@ def upload_to_frigate(jobs: list[dict]) -> None:
     ) as progress:
         upload_task = progress.add_task("[green]Uploading to Frigate", total=total_files)
 
-        for job in face_jobs:
+        for job in jobs:
             name = job["person"]["name"]
             # URL-encode the name for the API (handles spaces, special chars)
             encoded_name = quote(name, safe="")
             if " " in name:
                 progress.console.print(f"  ℹ️  URL-encoded name for Frigate API: '{name}' → '{encoded_name}'")
 
-            person_dir = os.path.join(Config.OUTPUT_DIR, name)
+            try:
+                person_dir = _safe_person_dir(Config.OUTPUT_DIR, name)
+            except ValueError as e:
+                logger.error(str(e))
+                continue
             if not os.path.isdir(person_dir):
                 progress.console.print(f"  [dim]⏭️  {name}: no output directory, skipping[/dim]")
                 continue
@@ -364,6 +378,8 @@ def upload_to_frigate(jobs: list[dict]) -> None:
             # Snapshot live Frigate files for post-upload reconciliation diff only.
             # effective_count is sourced from the tracker (mapped files) so that
             # manually-added Frigate files don't consume winnow's managed quota.
+            # Replacement targets also come exclusively from the tracker, so manually
+            # added files are never selected for deletion — only winnow-uploaded ones.
             _snapshot = (
                 all_frigate_files.get(name, []) if all_frigate_files is not None
                 else get_frigate_person_files(name)
@@ -438,14 +454,28 @@ def upload_to_frigate(jobs: list[dict]) -> None:
                         if _result is not None and (_result[0] or "").casefold() == name.casefold():
                             pre_fscore = _result[1]
 
-                # Ceiling check: skip if the existing training set already covers this
+                # Below-cap novelty gate: skip candidates already covered by the Frigate model,
-                # face condition well. Applies below cap only — at cap, replacement logic
+                # including conditions learned from manually-added images winnow can't track.
-                # drives the decision.
+                # pre_fscore is None on the first run (pre_run_count == 0 skips recognize_face
-                if not at_cap and Config.FRIGATE_SCORE_CEILING > 0 and pre_run_count > 0:
+                # above), so this block never fires on the first run without an extra guard.
-                    if pre_fscore is not None and pre_fscore > Config.FRIGATE_SCORE_CEILING:
+                if not at_cap and pre_fscore is not None:
+                    _ceiling = Config.FRIGATE_SCORE_CEILING
+                    if _ceiling is None:
+                        # Dynamic default: bar = most-redundant tracked file's Frigate score.
+                        # Falls back to uploading freely when no tracked scores exist yet.
+                        _bar = get_most_redundant_mapped_file(name)
+                        _skip = _bar is not None and pre_fscore > _bar[2]
+                        _bar_str = f"most redundant tracked {_bar[2]:.2f}" if _bar else ""
+                    elif _ceiling == 0.0:
+                        _skip = False  # explicitly disabled
+                        _bar_str = ""
+                    else:
+                        _skip = pre_fscore > _ceiling
+                        _bar_str = f"ceiling {_ceiling:.2f}"
+                    if _skip:
                         progress.console.print(
                             f"    [dim]⏭  {fname}: Frigate score {pre_fscore:.2f}"
-                            f" > ceiling {Config.FRIGATE_SCORE_CEILING:.2f}, already covered[/dim]"
+                            f" > {_bar_str}, already covered[/dim]"
                         )
                         progress.advance(upload_task)
                         continue
@@ -459,65 +489,45 @@ def upload_to_frigate(jobs: list[dict]) -> None:
                     using_fscore = person_has_fscores and Config.ENABLE_FRIGATE_SCORES
                     if using_fscore:
                         candidate_score = pre_fscore
-                        if candidate_score is None:
+                        get_target = get_most_redundant_mapped_file
-                            progress.console.print(
+                        score_label, better_note = "frigate", " (more novel)"
-                                f"    [dim]⏭  {fname}: Frigate recognize unavailable, skipping replacement[/dim]"
+                        no_score_msg = "Frigate recognize unavailable, skipping replacement"
-                            )
-                            progress.advance(upload_task)
-                            continue
-                        # Low score = more novel than the most redundant mapped file = replace
-                        target = get_most_redundant_mapped_file(name, exclude=failed_deletes)
-                        if target is None or candidate_score >= target[2]:
-                            target_score_str = f"{target[2]:.3f}" if target is not None else "N/A"
-                            progress.console.print(
-                                f"    [dim]⏭  {fname}: frigate {candidate_score:.3f} ≥ most redundant"
-                                f" {target_score_str}, not more novel[/dim]"
-                            )
-                            progress.advance(upload_task)
-                            continue
-                        target_frigate_file, _target_asset_id, target_score = target
-                        progress.console.print(
-                            f"    🔄 {fname}: frigate {candidate_score:.3f} < {target_score:.3f},"
-                            f" replacing {target_frigate_file} (more novel)"
-                        )
-                        if delete_frigate_person_files(name, [target_frigate_file]):
-                            remove_frigate_file(name, target_frigate_file)
-                            person_has_fscores = has_frigate_scores(name)
-                            effective_count -= 1
-                            # clear any blur-mode slot floor — Frigate uses a different score metric
-                            min_quality_score_for_slot = None
-                        else:
-                            logger.warning(f"Failed to delete {target_frigate_file} for {name}, skipping replacement")
-                            failed_deletes.add(target_frigate_file)
-                            progress.advance(upload_task)
-                            continue
                     else:
                         candidate_score = score_map.get(fname)
+                        get_target = get_lowest_quality_mapped_file
+                        score_label, better_note = "blur", ""
+                        no_score_msg = "no quality score, skipping replacement"
+
                     if candidate_score is None:
+                        progress.console.print(f"    [dim]⏭  {fname}: {no_score_msg}[/dim]")
+                        progress.advance(upload_task)
+                        continue
+
+                    target = get_target(name, exclude=failed_deletes)
+                    not_better = target is None or (
+                        candidate_score >= target[2] if using_fscore else candidate_score <= target[2]
+                    )
+                    if not_better:
+                        target_str = f"{target[2]:.3f}" if target is not None else "N/A"
+                        op = "<" if using_fscore else ">"
                         progress.console.print(
-                                f"    [dim]⏭  {fname}: no quality score, skipping replacement[/dim]"
+                            f"    [dim]⏭  {fname}: {score_label} {candidate_score:.3f}"
-                            )
+                            f" not {op} {target_str}, skipping[/dim]"
-                            progress.advance(upload_task)
-                            continue
-                        target = get_lowest_quality_mapped_file(name, exclude=failed_deletes)
-                        if target is None or candidate_score <= target[2]:
-                            target_score_str = f"{target[2]:.3f}" if target is not None else "N/A"
-                            progress.console.print(
-                                f"    [dim]⏭  {fname}: blur {candidate_score:.3f} ≤ worst"
-                                f" {target_score_str}, skipping[/dim]"
                         )
                         progress.advance(upload_task)
                         continue
+
                     target_frigate_file, _target_asset_id, target_score = target
+                    op = "<" if using_fscore else ">"
                     progress.console.print(
-                            f"    🔄 {fname}: blur {candidate_score:.3f} > {target_score:.3f},"
+                        f"    🔄 {fname}: {score_label} {candidate_score:.3f} {op} {target_score:.3f},"
-                            f" replacing {target_frigate_file}"
+                        f" replacing {target_frigate_file}{better_note}"
                     )
                     if delete_frigate_person_files(name, [target_frigate_file]):
                         remove_frigate_file(name, target_frigate_file)
                         person_has_fscores = has_frigate_scores(name)
                         effective_count -= 1
-                            min_quality_score_for_slot = score_map.get(fname)
+                        min_quality_score_for_slot = None if using_fscore else candidate_score
                     else:
                         logger.warning(f"Failed to delete {target_frigate_file} for {name}, skipping replacement")
                         failed_deletes.add(target_frigate_file)
@@ -564,13 +574,16 @@ def upload_to_frigate(jobs: list[dict]) -> None:
                             progress.console.print(
                                 f"    [red]✗ {fname}: HTTP {resp.status_code} (after {max_retries} attempts)[/red]"
                             )
+                            full_body = resp.text
                             try:
-                                error_detail = resp.json().get("message", resp.text[:100])
+                                error_detail = resp.json().get("message", full_body[:100])
-                                progress.console.print(f"      [dim]{error_detail}[/dim]")
                             except Exception:
-                                error_detail = resp.text[:100]
+                                error_detail = full_body[:100]
+                            if resp.status_code == 400:
                                 progress.console.print(f"      [dim]{error_detail}[/dim]")
-                            if resp.status_code == 400 and "face" in error_detail.lower():
+                            else:
+                                logger.debug(f"{fname} HTTP {resp.status_code}: {error_detail}")
+                            if resp.status_code == 400 and "face" in full_body.lower():
                                 asset_id = asset_map.get(fname)
                                 if asset_id:
                                     mark_rejected(asset_id, person_name=name)

winnow/image_processing.py

@@ -1,7 +1,8 @@
-"""Image processing functions for cropping faces and objects."""
+"""Image processing functions for cropping faces."""
 
 import logging
 import os
+import warnings
 
 import numpy as np
 from PIL import Image
@@ -10,9 +11,6 @@ from .config import Config
 
 logger = logging.getLogger(__name__)
 
-# Lazy singleton
-_yolo_model = None
-
 
 def _save_jpeg(img: Image.Image, path: str) -> None:
     if img.mode != "RGB":
@@ -20,17 +18,6 @@ def _save_jpeg(img: Image.Image, path: str) -> None:
     img.save(path, format="JPEG")
 
 
-def get_yolo_model():
-    """Singleton for YOLO model."""
-    global _yolo_model
-    if _yolo_model is None:
-        from ultralytics import YOLO
-
-        logger.info("Loading YOLOv9c model...")
-        _yolo_model = YOLO("yolov9c.pt")
-    return _yolo_model
-
-
 def align_face(img: Image.Image, landmarks: list[list[float]] | np.ndarray) -> Image.Image | None:
     """Align face using 5-point landmarks to standard ArcFace input format (112x112).
 
@@ -69,13 +56,15 @@ def process_face_mode(
     output_dir: str,
     count: int,
     min_width: int | None = None,
+    insightface_app=None,
 ) -> tuple[int, int] | None:
     """Crop face based on Immich metadata and save to output directory.
 
     Returns (width, height) of the saved crop, or None if no crop was saved.
-    If face alignment is enabled and landmarks are available, produces
+    When insightface_app is provided and ENABLE_FACE_ALIGNMENT is True,
-    an aligned 112x112 crop. Otherwise falls back to bounding box crop
+    re-detects the face in the Immich bbox region using InsightFace to get
-    with configurable margin.
+    precise landmarks for a proper 112x112 aligned crop. Falls back to
+    bounding box crop with configurable margin if alignment is unavailable.
     """
     min_width = min_width or Config.MIN_FACE_WIDTH
 
@@ -109,18 +98,53 @@ def process_face_mode(
         logger.debug(f"Face too small ({face_w:.1f}x{face_h:.1f})")
         return None
 
-    # Try face alignment if enabled and landmarks available
+    # Re-detect face with InsightFace for landmark-based alignment.
+    # Immich's /api/faces endpoint does not include landmarks, so the
+    # align_face fallback below never fires without this step.
+    if insightface_app is not None and Config.ENABLE_FACE_ALIGNMENT:
+        try:
+            # Expand the Immich bbox by 50% to give InsightFace enough context
+            # for detection and alignment, then search for the face nearest the
+            # centre of that region (handles group photos at the boundary).
+            pad_x, pad_y = face_w * 0.5, face_h * 0.5
+            search_box = (
+                max(0, x1 - pad_x),
+                max(0, y1 - pad_y),
+                min(img_w, x2 + pad_x),
+                min(img_h, y2 + pad_y),
+            )
+            search_crop = img.crop(search_box)
+            with warnings.catch_warnings():
+                warnings.filterwarnings("ignore", message=".*estimate.*is deprecated", category=FutureWarning)
+                detected = insightface_app.get(np.asarray(search_crop))
+            if detected:
+                cx, cy = search_crop.width / 2, search_crop.height / 2
+                best = min(
+                    detected,
+                    key=lambda f: abs((f.bbox[0] + f.bbox[2]) / 2 - cx)
+                    + abs((f.bbox[1] + f.bbox[3]) / 2 - cy),
+                )
+                kps = getattr(best, "kps", None)
+                if kps is not None and np.asarray(kps).shape == (5, 2):
+                    aligned = align_face(search_crop, kps)
+                    if aligned is not None:
+                        _save_jpeg(aligned, os.path.join(output_dir, f"{count}.jpg"))
+                        return aligned.size
+        except Exception as e:
+            logger.debug(f"InsightFace re-detection failed for {asset.get('id')}: {e}")
+
+    # Landmark alignment from Immich metadata (Immich does not currently
+    # expose landmarks, so this path is a future-proofing fallback)
     if Config.ENABLE_FACE_ALIGNMENT:
         landmarks = face_info.get("landmarks") or face_info.get("landmark")
         if landmarks:
-            # Scale landmarks
             scaled_landmarks = [[lm[0] * scale_x, lm[1] * scale_y] for lm in landmarks]
             aligned = align_face(img, scaled_landmarks)
             if aligned is not None:
                 _save_jpeg(aligned, os.path.join(output_dir, f"{count}.jpg"))
                 return aligned.size
 
-    # Fall back to bounding box crop with configurable margin
+    # Final fallback: bounding box crop with configurable margin
     margin = Config.FACE_MARGIN
     margin_x, margin_y = face_w * margin, face_h * margin
     crop_box = (
@@ -135,49 +159,4 @@ def process_face_mode(
     return face_crop.size
 
 
-def process_object_mode(
-    img: Image.Image,
-    config: dict,
-    output_dir: str,
-    count: int,
-) -> bool:
-    """Detect and crop objects using YOLO."""
-    try:
-        model = get_yolo_model()
-        target_class = config.get("object_class", "dog")
-        import torch
-
-        if os.getenv("FORCE_CPU", "").lower() in ("true", "1", "yes"):
-            device = "cpu"
-        elif hasattr(torch, "xpu") and torch.xpu.is_available():
-            device = "xpu"
-        else:
-            device = None  # YOLO auto-selects (CUDA/ROCm/CPU)
-
-        results = model(img, verbose=False, device=device)
-
-        found = False
-        class_idx = 0  # Sequential counter per target class (Issue #10)
-        for box in (box for r in results for box in r.boxes):
-            cls_id = int(box.cls[0])
-            conf = float(box.conf[0])
-            if 0 <= cls_id < len(model.names) and model.names[cls_id] == target_class and conf > 0.5:
-                x1, y1, x2, y2 = box.xyxy[0].tolist()
-                _save_jpeg(
-                    img.crop((x1, y1, x2, y2)),
-                    os.path.join(output_dir, f"{count}_{class_idx}.jpg"),
-                )
-                class_idx += 1
-                found = True
-
-        return found
-    except Exception as e:
-        logger.error(f"YOLO processing failed: {e}")
-        return False
-
-
-def process_full_mode(img: Image.Image, output_dir: str, count: int) -> bool:
-    """Save full image."""
-    _save_jpeg(img, os.path.join(output_dir, f"{count}.jpg"))
-    return True
 

winnow/immich_api.py

@@ -5,7 +5,6 @@ from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
 from io import BytesIO
 
-import numpy as np
 import requests
 from PIL import Image, ImageOps
 
@@ -14,13 +13,13 @@ from .config import Config, get_headers
 logger = logging.getLogger(__name__)
 
 MAX_PAGES = 1000  # Safety limit for pagination
+_MAX_ASSETS_PER_PERSON = 5000  # Stop fetching after this many — diversity pool is capped at 3000 anyway
 
 
 @dataclass
 class FaceData:
     """Pre-computed face data from Immich."""
 
-    embedding: np.ndarray | None
     bbox: tuple[float, float, float, float]  # (x1, y1, x2, y2)
     confidence: float | None
     image_width: int
@@ -45,6 +44,26 @@ def get_people() -> list[dict]:
         return []
 
 
+def merge_people(survivor_id: str, merge_ids: list[str]) -> bool:
+    """Merge duplicate people into survivor via Immich's merge endpoint.
+
+    The survivor (identified by survivor_id) absorbs all faces and assets
+    from the people in merge_ids, which are then removed from Immich.
+    """
+    try:
+        resp = requests.put(
+            f"{Config.IMMICH_URL}/api/people/{survivor_id}/merge",
+            headers={**get_headers(), "Content-Type": "application/json"},
+            json={"ids": merge_ids},
+            timeout=30,
+        )
+        resp.raise_for_status()
+        return True
+    except requests.RequestException as e:
+        logger.error(f"Failed to merge people into {survivor_id}: {e}")
+        return False
+
+
 def fetch_all_assets(person: dict) -> list[dict]:
     """Fetch all assets for a person with pagination."""
     name = person.get("name", "Unknown")
@@ -75,10 +94,10 @@ def fetch_all_assets(person: dict) -> list[dict]:
             if not page_assets:
                 break
 
-            assets.extend(page_assets)
+            assets.extend(a for a in page_assets if isinstance(a, dict))
             logger.debug(f"Fetched page {page}, total: {len(assets)}")
 
-            if len(page_assets) < page_size:
+            if len(page_assets) < page_size or len(assets) >= _MAX_ASSETS_PER_PERSON:
                 break
 
         except (requests.RequestException, ValueError) as e:
@@ -89,7 +108,7 @@ def fetch_all_assets(person: dict) -> list[dict]:
 
 
 def fetch_face_data(asset_id: str, person_id: str | None = None) -> FaceData | None:
-    """Fetch pre-computed face data (embedding, bbox, confidence) from Immich.
+    """Fetch pre-computed face data (bbox, confidence) from Immich.
 
     Queries GET /api/faces?id={asset_id} to retrieve face detection results
     that Immich already computed using InsightFace Buffalo_L.
@@ -99,7 +118,7 @@ def fetch_face_data(asset_id: str, person_id: str | None = None) -> FaceData | N
         person_id: Optional person ID to match the specific face
 
     Returns:
-        FaceData with embedding, bbox, and confidence, or None if unavailable
+        FaceData with bbox and confidence, or None if unavailable
     """
     try:
         resp = requests.get(
@@ -127,12 +146,6 @@ def fetch_face_data(asset_id: str, person_id: str | None = None) -> FaceData | N
         if face is None:
             face = faces[0]  # Fall back to first/largest face
 
-        # Extract embedding if available
-        embedding = None
-        if "embedding" in face:
-            embedding = np.array(face["embedding"], dtype=np.float32)
-
-        # Extract bounding box
         bbox = (
             face.get("boundingBoxX1", 0),
             face.get("boundingBoxY1", 0),
@@ -142,7 +155,6 @@ def fetch_face_data(asset_id: str, person_id: str | None = None) -> FaceData | N
 
         score = face.get("score")
         return FaceData(
-            embedding=embedding,
             bbox=bbox,
             confidence=score if score is not None else face.get("confidence"),
             image_width=face.get("imageWidth", 0),

winnow/jobs.py

@@ -20,18 +20,16 @@ logger = logging.getLogger(__name__)
 
 # Strategy presets: (limit, mode_name)
 STRATEGY_PRESETS = {
-    "1": ("auto", "Auto Diversity"),
+    "1": ("auto", "Adaptive Diversity"),
     "2": (30, "Standard (30)"),
     "3": (100, "Broad (100)"),
 }
 
 
-def _get_strategy_choice(has_embedding: bool, entity_type: str) -> tuple[int | str, str]:
+def _get_strategy_choice(has_embedding: bool) -> tuple[int | str, str]:
     """Prompt user for training strategy and return (limit, selection_mode)."""
-    model_name = "InsightFace" if entity_type == "face" else "SigLIP"
-
     if has_embedding:
-        rprint("  [bold]1.[/bold] Auto (Objective Diversity) [green][Recommended][/green]")
+        rprint("  [bold]1.[/bold] Adaptive Diversity [green][Recommended][/green]")
         rprint("     [dim]• Dynamically selects images until redundancy starts[/dim]")
         rprint("  [bold]2.[/bold] Standard (30 images)")
         rprint("  [bold]3.[/bold] Broad (100 images)")
@@ -51,7 +49,7 @@ def _get_strategy_choice(has_embedding: bool, entity_type: str) -> tuple[int | s
         return 30, "smart"
 
     # Fallback when embedding model not available
-    rprint(f"  [yellow]Note: {model_name} not available. Using Time Spread.[/yellow]")
+    rprint("  [yellow]Note: InsightFace not available. Using Time Spread.[/yellow]")
     rprint("  [bold]1.[/bold] Standard (30 images) [green][Recommended][/green]")
     rprint("  [bold]2.[/bold] Broad (100 images)")
     rprint("  [bold]3.[/bold] Custom Count")
@@ -77,7 +75,8 @@ def _resolve_strategy(strategy: str, has_embedding: bool) -> tuple[int | str, st
         return int(custom_limit), "smart"
 
     strategy_map = {
-        "auto": ("auto", "smart"),
+        "adaptive": ("auto", "smart"),
+        "auto": ("auto", "smart"),  # legacy alias for adaptive
         "standard": (30, "smart"),
         "broad": (100, "smart"),
     }
@@ -85,15 +84,13 @@ def _resolve_strategy(strategy: str, has_embedding: bool) -> tuple[int | str, st
 
 
 def _perform_selection(
-    assets: list, limit: int | str, name: str, selection_mode: str, entity_type: str, person_id: str | None = None
+    assets: list, limit: int | str, name: str, selection_mode: str, person_id: str | None = None
 ) -> list:
     """Run diversity selection with progress display."""
     if selection_mode == "smart":
-        model_display = "InsightFace (face embeddings)" if entity_type == "face" else "SigLIP (visual embeddings)"
+        rprint("\n[cyan]Using InsightFace (face embeddings) for diversity analysis...[/cyan]")
-        rprint(f"\n[cyan]Using {model_display} for diversity analysis...[/cyan]")
 
-        # Pre-load model explicitly (separate from availability check)
+        load_embedding_model()
-        load_embedding_model(entity_type)
 
         with Progress(
             SpinnerColumn(),
@@ -108,7 +105,6 @@ def _perform_selection(
                 limit,
                 name,
                 selection_mode=selection_mode,
-                entity_type=entity_type,
                 person_id=person_id,
                 progress_callback=lambda c, t: progress.update(task, completed=c, total=t),
             )
@@ -119,9 +115,7 @@ def _perform_selection(
 
     rprint(f"\n[cyan]Using time-spread selection for {limit} images...[/cyan]")
     with console.status(f"[bold]Selecting {limit} images evenly distributed over time...[/bold]"):
-        selected = select_diverse_assets(
+        selected = select_diverse_assets(assets, limit, name, selection_mode="time", person_id=person_id)
-            assets, limit, name, selection_mode="time", entity_type=entity_type, person_id=person_id
-        )
     rprint(f"  [green]Selected {len(selected)} images using time spread.[/green]")
     return selected
 
@@ -131,22 +125,12 @@ def _configure_person(person: dict, people: list[dict]) -> dict | None:
     name = person["name"]
     console.print(f"\nSelected: [bold green]{name}[/bold green]")
 
-    # Select training mode
+    config = {"name": name, "quality_replacement": Config.QUALITY_REPLACEMENT}
-    rprint("\n[bold cyan]Training Mode:[/bold cyan]")
-    rprint("  [bold]1.[/bold] Face (Frigate Face Recognition)")
-    rprint("  [bold]2.[/bold] Object (Frigate Object Classification)")
-
-    mode_choice = Prompt.ask("Choice", choices=["1", "2"], default="1")
-    entity_type = "face" if mode_choice == "1" else "object"
-
-    config = {"name": name, "mode": entity_type, "quality_replacement": Config.QUALITY_REPLACEMENT}
-    if entity_type == "object":
-        config["object_class"] = Prompt.ask("Enter Object Class (e.g. dog, cat, car)", default="dog")
 
     # Fetch and filter assets
     years = IntPrompt.ask("Filter images older than (years)", default=Config.YEARS_FILTER)
 
-    console.print(f"Scanning for {name} ({entity_type})...")
+    console.print(f"Scanning for {name}...")
     with console.status("[bold green]Fetching assets...[/bold green]"):
         all_assets = fetch_all_assets(person)
         recent_assets = filter_recent_assets(all_assets, years=years)
@@ -170,17 +154,15 @@ def _configure_person(person: dict, people: list[dict]) -> dict | None:
         return None
 
     # Strategy selection
-    has_embedding = is_embedding_available(entity_type)
+    has_embedding = is_embedding_available()
     rprint(f"\n[bold cyan]Select Training Strategy for {name}:[/bold cyan]")
 
-    limit, selection_mode = _get_strategy_choice(has_embedding, entity_type)
+    limit, selection_mode = _get_strategy_choice(has_embedding)
     if selection_mode == "skip":
         return None
 
     # Perform selection
-    selected_assets = _perform_selection(
+    selected_assets = _perform_selection(recent_assets, limit, name, selection_mode, person_id=person["id"])
-        recent_assets, limit, name, selection_mode, entity_type, person_id=person["id"]
-    )
 
     rprint(f"  [green]Queued {len(selected_assets)} images for {name}.[/green]")
     return {"person": person, "assets": selected_assets, "limit": len(selected_assets), "config": config}
@@ -230,7 +212,6 @@ def auto_configure(people: list[dict]) -> list[dict]:
         rprint("[red]No people found with names in Immich.[/red]")
         return []
 
-    mode = os.environ.get("TRAINING_MODE", "face")
     strategy = os.environ.get("STRATEGY", "auto")
     skip = os.environ.get("SKIP_PEOPLE", "").split(",") if os.environ.get("SKIP_PEOPLE") else []
     only = os.environ.get("ONLY_PEOPLE", "").split(",") if os.environ.get("ONLY_PEOPLE") else []
@@ -259,11 +240,7 @@ def auto_configure(people: list[dict]) -> list[dict]:
     jobs = []
     for person in valid_people:
         name = person["name"]
-        entity_type = mode
+        config = {"name": name}
-
-        config = {"name": name, "mode": entity_type}
-        if entity_type == "object":
-            config["object_class"] = os.environ.get("OBJECT_CLASS", "dog")
 
         all_assets = fetch_all_assets(person)
         recent_assets = filter_recent_assets(all_assets, years=Config.YEARS_FILTER)
@@ -306,7 +283,7 @@ def auto_configure(people: list[dict]) -> list[dict]:
 
         config["quality_replacement"] = quality_replacement_only or Config.QUALITY_REPLACEMENT
 
-        has_embedding = is_embedding_available(entity_type)
+        has_embedding = is_embedding_available()
         limit, selection_mode = _resolve_strategy(strategy, has_embedding)
 
         # Cap selection to remaining capacity (no cap when replacement-only — executor
@@ -322,9 +299,7 @@ def auto_configure(people: list[dict]) -> list[dict]:
         if selection_mode == "skip":
             continue
 
-        selected_assets = _perform_selection(
+        selected_assets = _perform_selection(recent_assets, limit, name, selection_mode, person_id=person["id"])
-            recent_assets, limit, name, selection_mode, entity_type, person_id=person["id"]
-        )
         if auto_cap is not None:
             selected_assets = selected_assets[:auto_cap]
 
@@ -339,20 +314,17 @@ def _show_preview(jobs: list[dict]) -> None:
     """Show a summary table of all queued jobs before execution."""
     table = Table(title="📋 Training Job Preview", show_header=True, header_style="bold cyan")
     table.add_column("Person", style="bold")
-    table.add_column("Mode", style="dim")
     table.add_column("Images", justify="right")
     table.add_column("Date Range", style="dim")
 
     for job in jobs:
         name = job["person"]["name"]
-        mode = job["config"].get("mode", "face")
         count = str(job["limit"])
 
-        # Date range
         dates = sorted(a.get("fileCreatedAt", "")[:10] for a in job["assets"] if a.get("fileCreatedAt"))
         date_range = f"{dates[0]} → {dates[-1]}" if len(dates) >= 2 else (dates[0] if dates else "—")
 
-        table.add_row(name, mode, count, date_range)
+        table.add_row(name, count, date_range)
 
     console.print()
     console.print(table)

winnow/log_config.py

@@ -13,12 +13,9 @@ console = Console()
 NOISY_LOGGERS = (
     "urllib3",
     "PIL",
-    "ultralytics",
     "insightface",
     "onnxruntime",
     "matplotlib",
-    "transformers",
-    "torch",
 )
 
 
@@ -51,7 +48,6 @@ def setup_logging(verbose: bool = False) -> logging.Logger:
 
     # Suppress Python warnings from ML libraries
     warnings.filterwarnings("ignore", category=UserWarning, module="onnxruntime")
-    warnings.filterwarnings("ignore", category=FutureWarning, module="transformers")
 
     return root
 

winnow/upload_tracker.py

@@ -39,6 +39,11 @@ logger = logging.getLogger(__name__)
 UPLOAD_TRACKER_FILE = "frigate_uploaded_ids.json"
 REJECT_TRACKER_FILE = "frigate_rejected_ids.json"
 
+# Write-through in-memory cache keyed by the resolved file path.
+# Reduces per-call JSON reads from O(calls) to O(1) after the first load.
+# Keyed by full path so tests with isolated tmp dirs never share entries.
+_cache: dict[str, dict] = {}
+
 
 def _tracker_path(filename: str) -> Path:
     try:
@@ -50,18 +55,23 @@ def _tracker_path(filename: str) -> Path:
 
 def _load(filename: str) -> dict:
     path = _tracker_path(filename)
-    if not path.exists():
+    key = str(path)
-        return {}
+    if key in _cache:
+        return _cache[key]
+    data: dict = {}
+    if path.exists():
         try:
             with open(path) as f:
-            return json.load(f)
+                data = json.load(f)
         except (json.JSONDecodeError, OSError) as e:
             logger.warning(f"Could not load tracker {filename}: {e}")
-        return {}
+    _cache[key] = data
+    return data
 
 
 def _save(filename: str, data: dict) -> None:
     path = _tracker_path(filename)
+    _cache[str(path)] = data  # keep cache consistent with what we write
     path.parent.mkdir(parents=True, exist_ok=True)
     with open(path, "w") as f:
         json.dump(data, f, indent=2)
@@ -161,6 +171,19 @@ def record_frigate_file(person_name: str, frigate_filename: str, asset_id: str)
     logger.debug(f"Mapped Frigate file {frigate_filename} → {asset_id} ({person_name})")
 
 
+def record_frigate_files_batch(person_name: str, mappings: dict[str, str]) -> None:
+    """Record multiple Frigate filename → asset_id mappings in a single load/save."""
+    if not mappings:
+        return
+    data = _load(UPLOAD_TRACKER_FILE)
+    by_person = data.setdefault("by_person", {})
+    entry = _migrate_entry(by_person.get(person_name, {}))
+    entry["frigate_files"].update(mappings)
+    by_person[person_name] = entry
+    _save(UPLOAD_TRACKER_FILE, data)
+    logger.debug(f"Batch-mapped {len(mappings)} Frigate file(s) for {person_name}")
+
+
 def remove_frigate_file(person_name: str, frigate_filename: str) -> None:
     """Remove a Frigate filename from the mapping after it has been deleted.