feat: Implement advanced diversity selection algorithms, comprehensive quality filtering, and caching for improved image curation.

.gitignore

@@ -5,6 +5,7 @@ yolov9c.pt
 .insightface/
 .huggingface/
 .cache/huggingface
+.if_cache/
 
 .immich_config.json
 # Python-generated files

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Sebastian
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -11,31 +11,6 @@
 
 </div>
 
-
----
-
-> [!WARNING]  
-> This is 100% vibe-coded.
-
-
-
-
-
-## ⚡ Why This Tool?
-
-> **"Diversity matters far more than volume."** — *Frigate Developer Tips*
-
-Training AI models on "bulk" data is often harmful. If you feed the model 50 images from the same 10-second video clip, it learns to recognize the *lighting and background*, not the actual *face* or *object*.
-
-`if-curator` solves this using **AI-powered diversity selection**:
-
-| Mode | Embedding Model | Algorithm |
-| :--- | :--- | :--- |
-| **👤 Face** | InsightFace (ArcFace) | Farthest Point Sampling |
-| **🐶 Object** | SigLIP (Vision Transformer) | Farthest Point Sampling |
-
-Both use the same **Farthest Point Sampling (FPS)** algorithm that mathematically selects images until redundancy starts, ensuring optimal diversity whether that's 20 or 150 images.
-
 > [!WARNING]
 > **Regarding Object Classification**
 >
@@ -44,25 +19,67 @@ Both use the same **Farthest Point Sampling (FPS)** algorithm that mathematicall
 
 ---
 
+## ⚡ Why This Tool?
+
+> **"Diversity matters far more than volume."** — *Frigate Developer Tips*
+
+Training AI models on "bulk" data is often harmful. If you feed the model 50 images from the same 10-second video clip, it learns to recognize the *lighting and background*, not the actual *face* or *object*.
+
+`if-curator` solves this using **AI-powered diversity selection** and **quality filtering**:
+
+| Mode | Embedding Model | Algorithm |
+| :--- | :--- | :--- |
+| **👤 Face** | InsightFace (ArcFace) | K-Medoids Clustering + FPS + Hard Example Weighting |
+| **🐶 Object** | SigLIP (Vision Transformer) | K-Medoids Clustering + FPS |
+
+The pipeline uses **K-Medoids clustering** to guarantee coverage of every distinct "look", then fills the remaining budget with **Farthest Point Sampling (FPS)** biased toward **hard examples** (unusual angles, partial occlusions). An **adaptive threshold** stops selection automatically when adding more images becomes redundant.
+
+---
+
 ## ✨ Features
 
-### 🎯 Unified Selection Strategies
+### 🎯 Smart Selection
-Both Face and Object modes offer the same powerful options:
+- **Auto Diversity [Recommended]**: Clusters images by visual similarity, selects representatives from each cluster, then fills with maximally-diverse picks until redundancy starts (capped at 80)
-- **Auto (Objective Diversity) [Recommended]**: Dynamically selects images until redundancy starts
 - **Standard (30 images)**: Balanced set using Smart Diversity
 - **Broad (100 images)**: Extensive set using Smart Diversity
 - **Custom Count**: You choose the limit
 
+### � Quality Filtering
+Bad training data hurts ArcFace models. Images are automatically rejected if they are:
+- **Blurry** — Laplacian variance below threshold
+- **Grayscale / IR** — ArcFace is trained on color images only
+- **Over/Underexposed** — Washed-out or too dark to use
+- **Low confidence** — Partial or occluded face detections
+- **Too small** — Faces under 100px (configurable) lack features
+
 ### 👤 Face Recognition Prep
-- Uses **InsightFace** (ArcFace/Buffalo_L) embeddings
+- Uses **InsightFace** (ArcFace/Buffalo_L) embeddings on **face crops** (not full images — avoids wrong-face in group photos)
-- Extracts faces using Immich's metadata
+- **Hard example prioritization** — unusual angles, sunglasses, and low-confidence detections are biased for selection
-- **Auto-Diversity** picks the optimal set size based on visual distinctness
+- **Face alignment** via InsightFace landmarks (standard 112×112 ArcFace input)
+- Downloads **full-resolution** originals for final crops (falls back to JPEG preview for HEIC/RAW)
+- Configurable crop margin (default 15%)
 
 ### 📦 Object/State Classification Prep
 - Uses **SigLIP** (Vision Transformer) embeddings for semantic diversity
 - **YOLOv9c** to detect and crop specific objects (dogs, cars, etc.)
 - Captures variation in poses, lighting, and backgrounds
-- *Note: As mentioned, Frigate upload is pending support.*
+
+### ⚡ Performance
+- **Concurrent thumbnail downloads** (8 parallel workers)
+- **Batch-capable** SigLIP embeddings for GPU efficiency
+- Optional **disk-based embedding cache** for faster re-runs
+- **Multi-person batch mode** — process multiple people in one session
+
+### 📋 Preview Before Download
+After selection, a summary table shows what will be processed:
+```
+                📋 Training Job Preview
+┏━━━━━━━━━━━┳━━━━━━┳━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━┓
+┃ Person    ┃ Mode ┃ Images ┃ Date Range              ┃
+┡━━━━━━━━━━━╇━━━━━━╇━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━┩
+│ Sebastian │ face │     80 │ 2021-04-03 → 2026-02-18 │
+└───────────┴──────┴────────┴─────────────────────────┘
+```
 
 ---
 
@@ -93,22 +110,25 @@ uv sync --extra gpu
 
 ## 💻 Usage
 
-Run the command-line interface:
-
 ```bash
-uv run -m if_curator
+uv run if-curator
 ```
 
 ### Interactive Flow
 The tool will guide you through:
-1.  **Select Person/Subject**: Choose from your Immich people.
+1.  **Select Person** — Choose from your Immich people (supports multi-person batch)
-2.  **Training Mode**: Face (Recognition) or Object (Classification).
+2.  **Training Mode** — Face (Recognition) or Object (Classification)
-3.  **Strategy**: Auto, Standard, Broad, etc.
+3.  **Strategy** — Auto, Standard, Broad, or Custom
+4.  **Preview** — Review the selection summary before downloading
+5.  **Execute** — Downloads and processes images with progress tracking
 
 ```text
-Using SigLIP (visual embeddings) for diversity analysis...
+Using InsightFace (face embeddings) for diversity analysis...
-Computing embeddings for 69 images... ████████████████ 100%
+Quality filtering removed 76 images.
-Auto-diversity selected 38 optimally diverse images.
+Adaptive threshold: 0.1721 (median_dist=0.8605, fraction=0.2)
+Clustering 223 embeddings into 20 groups (K-Medoids)...
+Selected 20 cluster medoids as initial picks.
+Selection complete: 80 images (0 hard examples with confidence < 0.85).
 ```
 
 ---
@@ -117,18 +137,55 @@ Auto-diversity selected 38 optimally diverse images.
 
 The tool prompts for your Immich URL and API Key on the first run and saves them to `.immich_config.json`.
 
-| Variable | Description |
+### Environment Variables
-| :--- | :--- |
+
-| `IMMICH_URL` | Full URL to Immich (e.g. `http://192.168.1.10:2283`) |
+| Variable | Default | Description |
-| `API_KEY` | Your Immich API Key |
+| :--- | :--- | :--- |
-| `FORCE_CPU` | Set to `true` to disable GPU acceleration |
+| `IMMICH_URL` | — | Full URL to Immich (e.g. `http://192.168.1.10:2283`) |
+| `API_KEY` | — | Your Immich API Key |
+| `FORCE_CPU` | `false` | Disable GPU acceleration |
+| `MIN_FACE_WIDTH` | `100` | Minimum face crop size (pixels) |
+| `BLUR_THRESHOLD` | `100.0` | Laplacian variance threshold for blur detection |
+| `MIN_CONFIDENCE` | `0.7` | Minimum Immich detection confidence |
+| `MAX_AUTO_IMAGES` | `80` | Safety cap for auto-diversity mode |
+| `FACE_MARGIN` | `0.15` | Crop margin around face (fraction) |
+| `USE_FULL_RESOLUTION` | `true` | Download originals for final crops |
+| `ENABLE_FACE_ALIGNMENT` | `true` | Align faces to ArcFace 112×112 format |
+| `ENABLE_CACHE` | `false` | Cache embeddings to disk for faster re-runs |
+| `CACHE_DIR` | `.if_cache` | Directory for embedding cache |
 
 ---
 
 ## 🧠 Technical Details
 
-- **InsightFace**: Face detection and embedding (ArcFace)
+### Models
-- **SigLIP**: Visual embeddings via `transformers` (OpenAI CLIP alternative)
+- **InsightFace (Buffalo_L)** — Face detection and embedding (ArcFace, 512-d)
-- **YOLOv9c**: State-of-the-art object detection for cropping
+- **SigLIP** — Visual embeddings via `transformers` (google/siglip-base-patch16-224, 768-d)
-- **Rich**: Beautiful terminal UI
+- **YOLOv9c** — Object detection for cropping
 
+### Algorithms
+- **K-Medoids Clustering** — Groups embeddings into k clusters using cosine distance, selecting actual data points (medoids) as cluster centers. Guarantees one representative from every distinct "look"
+- **Farthest Point Sampling** — After medoid selection, fills remaining budget by iteratively selecting the most distant point from the current set
+- **Hard Example Weighting** — Candidates with detection confidence < 0.85 get a 1.2–1.5× distance boost, biasing selection toward challenging images (unusual angles, occlusions)
+- **Adaptive Auto-Threshold** — Computed as 20% of the median pairwise cosine distance; stops when the next-best image is too similar
+- **Quality Filtering** — Blur (Laplacian), grayscale/IR (channel comparison), exposure (histogram), confidence (Immich metadata)
+- **Face Crop Embedding** — Extracts the target person's face (using Immich bbox) before embedding, preventing wrong-face selection in group photos
+
+### Architecture
+```
+Immich API ─► Fetch Assets by Person ─► Time Filter
+                                            │
+                                  Concurrent Thumbnail Download (8 workers)
+                                            │
+                                  Quality Filtering (blur, IR, exposure...)
+                                            │
+                                  Face Crop Extraction (bbox from Immich metadata)
+                                            │
+                                  Compute Embeddings (InsightFace / SigLIP)
+                                            │
+                                  K-Medoids Clustering → FPS + Hard Example Weighting
+                                            │
+                                  Preview Summary Table
+                                            │
+                                  Download Full-Res ─► Face Alignment ─► Save
+```

if_curator/__init__.py

@@ -0,0 +1,8 @@
+"""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
+Classification models.
+"""
+
+__version__ = "0.1.0"

if_curator/cache.py

@@ -0,0 +1,87 @@
+"""Disk-based embedding cache.
+
+Caches embeddings keyed by (asset_id, model_version) to avoid
+recomputing on reruns. Uses numpy binary format for fast I/O.
+"""
+
+import hashlib
+import logging
+import os
+
+import numpy as np
+
+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",
+}
+
+
+class EmbeddingCache:
+    """Simple disk-based embedding cache.
+
+    Embeddings are stored as .npy files in a flat directory,
+    keyed by a hash of (asset_id, model_version).
+    """
+
+    def __init__(self, cache_dir: str = ".if_cache") -> None:
+        self.cache_dir = cache_dir
+        self._ensured = False
+
+    def _ensure_dir(self) -> None:
+        if not self._ensured:
+            os.makedirs(self.cache_dir, exist_ok=True)
+            self._ensured = True
+
+    @staticmethod
+    def _key(asset_id: str, model: str) -> str:
+        version = MODEL_VERSIONS.get(model, model)
+        raw = f"{asset_id}:{version}"
+        return hashlib.sha256(raw.encode()).hexdigest()[:16]
+
+    def _path(self, asset_id: str, model: str) -> str:
+        return os.path.join(self.cache_dir, f"{self._key(asset_id, model)}.npy")
+
+    def get(self, asset_id: str, model: str = "insightface") -> np.ndarray | None:
+        """Retrieve cached embedding, or None if not cached."""
+        path = self._path(asset_id, model)
+        if os.path.exists(path):
+            try:
+                return np.load(path)
+            except Exception:
+                return None
+        return None
+
+    def put(self, asset_id: str, embedding: np.ndarray, model: str = "insightface") -> None:
+        """Store an embedding in the cache."""
+        self._ensure_dir()
+        try:
+            np.save(self._path(asset_id, model), embedding)
+        except Exception as e:
+            logger.debug(f"Cache write failed for {asset_id}: {e}")
+
+    def clear(self) -> None:
+        """Delete all cached embeddings."""
+        if not os.path.isdir(self.cache_dir):
+            return
+        count = 0
+        for f in os.listdir(self.cache_dir):
+            if f.endswith(".npy"):
+                os.remove(os.path.join(self.cache_dir, f))
+                count += 1
+        logger.info(f"Cleared {count} cached embeddings.")
+
+
+# Singleton instance
+_cache: EmbeddingCache | None = None
+
+
+def get_cache(cache_dir: str = ".if_cache") -> EmbeddingCache:
+    """Get or create the singleton cache instance."""
+    global _cache
+    if _cache is None:
+        _cache = EmbeddingCache(cache_dir)
+    return _cache

if_curator/cli.py

@@ -9,13 +9,15 @@ from PIL import Image
 from rich import print as rprint
 from rich.progress import BarColumn, Progress, SpinnerColumn, TaskProgressColumn, TextColumn
 from rich.prompt import Confirm, IntPrompt, Prompt
+from rich.table import Table
 
 from .config import Config, ConfigManager
 from .diversity import select_diverse_assets
 from .embeddings import is_embedding_available
 from .image_processing import process_face_mode, process_full_mode, process_object_mode
-from .immich_api import fetch_all_assets, filter_recent_assets, get_people
+from .immich_api import fetch_all_assets, fetch_full_image, filter_recent_assets, get_people
 from .logging import console, setup_logging
+
 logger = logging.getLogger(__name__)
 
 # Strategy presets: (limit, mode_name)
@@ -62,23 +64,9 @@ def _get_strategy_choice(has_embedding: bool, entity_type: str) -> tuple[int | s
     return limits.get(choice, 0), "time" if choice != "4" else "skip"
 
 
-def interactive_configure(people: list[dict]) -> list[dict]:
+def _configure_person(person: dict, people: list[dict]) -> dict | None:
-    """Interactive phase: select person, mode, and configure training strategy."""
+    """Configure training for a single person. Returns job dict or None."""
-    valid_people = sorted([p for p in people if p.get("name")], key=lambda x: x["name"])
-
-    if not valid_people:
-        rprint("[red]No people found with names in Immich.[/red]")
-        return []
-
-    # Select person
-    console.print("\n[bold cyan]Select Person to Train:[/bold cyan]")
-    for idx, p in enumerate(valid_people, 1):
-        console.print(f"  [bold]{idx}.[/bold] {p['name']}")
-
-    p_choice = IntPrompt.ask("Enter Number", choices=[str(i) for i in range(1, len(valid_people) + 1)])
-    person = valid_people[p_choice - 1]
     name = person["name"]
-
     console.print(f"\nSelected: [bold green]{name}[/bold green]")
 
     # Select training mode
@@ -105,7 +93,7 @@ def interactive_configure(people: list[dict]) -> list[dict]:
 
     if not recent_assets:
         rprint("  [dim]Skipping (0 recent images).[/dim]")
-        return []
+        return None
 
     # Strategy selection
     has_embedding = is_embedding_available(entity_type)
@@ -113,18 +101,52 @@ def interactive_configure(people: list[dict]) -> list[dict]:
 
     limit, selection_mode = _get_strategy_choice(has_embedding, entity_type)
     if selection_mode == "skip":
-        return []
+        return None
 
     # Perform selection
     selected_assets = _perform_selection(recent_assets, limit, name, selection_mode, entity_type)
 
     rprint(f"  [green]Queued {len(selected_assets)} images for {name}.[/green]")
-    return [{"person": person, "assets": selected_assets, "limit": len(selected_assets), "config": config}]
+    return {"person": person, "assets": selected_assets, "limit": len(selected_assets), "config": config}
 
 
-def _perform_selection(
+def interactive_configure(people: list[dict]) -> list[dict]:
-    assets: list, limit: int | str, name: str, selection_mode: str, entity_type: str
+    """Interactive phase: select person(s), mode, and configure training strategy.
-) -> list:
+
+    Supports multi-person batch mode — after configuring one person,
+    prompts to add another.
+    """
+    valid_people = sorted([p for p in people if p.get("name")], key=lambda x: x["name"])
+
+    if not valid_people:
+        rprint("[red]No people found with names in Immich.[/red]")
+        return []
+
+    jobs = []
+
+    while True:
+        # Select person
+        console.print("\n[bold cyan]Select Person to Train:[/bold cyan]")
+        for idx, p in enumerate(valid_people, 1):
+            # Mark already-queued people
+            marker = " [dim](queued)[/dim]" if any(j["person"]["id"] == p["id"] for j in jobs) else ""
+            console.print(f"  [bold]{idx}.[/bold] {p['name']}{marker}")
+
+        p_choice = IntPrompt.ask("Enter Number", choices=[str(i) for i in range(1, len(valid_people) + 1)])
+        person = valid_people[p_choice - 1]
+
+        job = _configure_person(person, valid_people)
+        if job:
+            jobs.append(job)
+
+        # Multi-person: ask to add another
+        if not Confirm.ask("\nAdd another person?", default=False):
+            break
+
+    return jobs
+
+
+def _perform_selection(assets: list, limit: int | str, name: str, selection_mode: str, entity_type: str) -> list:
     """Run diversity selection with progress display."""
     if selection_mode == "smart":
         model_display = "InsightFace (face embeddings)" if entity_type == "face" else "SigLIP (visual embeddings)"
@@ -134,12 +156,17 @@ def _perform_selection(
         is_embedding_available(entity_type)
 
         with Progress(
-            SpinnerColumn(), TextColumn("[progress.description]{task.description}"),
+            SpinnerColumn(),
-            BarColumn(), TaskProgressColumn(), console=console,
+            TextColumn("[progress.description]{task.description}"),
+            BarColumn(),
+            TaskProgressColumn(),
+            console=console,
         ) as progress:
             task = progress.add_task(f"[cyan]Computing embeddings for {len(assets)} images...", total=None)
             selected = select_diverse_assets(
-                assets, limit, name,
+                assets,
+                limit,
+                name,
                 selection_mode=selection_mode,
                 entity_type=entity_type,
                 progress_callback=lambda c, t: progress.update(task, completed=c, total=t),
@@ -156,6 +183,30 @@ def _perform_selection(
     return selected
 
 
+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)
+
+    console.print()
+    console.print(table)
+    console.print()
+
+
 def execute_jobs(jobs: list[dict]) -> None:
     """Download and process images for all jobs."""
     if not jobs:
@@ -163,9 +214,14 @@ def execute_jobs(jobs: list[dict]) -> None:
 
     console.rule("[bold blue]Execution Phase")
 
+    use_full_res = Config.USE_FULL_RESOLUTION
+
     with Progress(
-        SpinnerColumn(), TextColumn("[progress.description]{task.description}"),
+        SpinnerColumn(),
-        BarColumn(), TaskProgressColumn(), console=console,
+        TextColumn("[progress.description]{task.description}"),
+        BarColumn(),
+        TaskProgressColumn(),
+        console=console,
     ) as progress:
         grand_total = sum(j["limit"] for j in jobs)
         overall_task = progress.add_task("[green]Overall Progress", total=grand_total)
@@ -181,17 +237,22 @@ def execute_jobs(jobs: list[dict]) -> None:
             count = 0
             for asset in assets:
                 try:
+                    # Use full-resolution for final output when configured
+                    if use_full_res:
+                        img = fetch_full_image(asset["id"])
+                    else:
                         resp = requests.get(
                             f"{Config.IMMICH_URL}/api/assets/{asset['id']}/thumbnail?size=preview&format=JPEG",
                             headers={"x-api-key": Config.API_KEY, "Accept": "application/json"},
                             timeout=30,
                         )
-                    if not resp.ok:
+                        img = Image.open(BytesIO(resp.content)) if resp.ok else None
+
+                    if img is None:
                         progress.console.print(f"[red]Failed download {asset['id']}[/red]")
                     else:
-                        img = Image.open(BytesIO(resp.content))
                         saved = (
-                            process_face_mode(img, asset, person, person_dir, count, min_width=Config.MIN_FACE_WIDTH)
+                            process_face_mode(img, asset, person, person_dir, count)
                             if mode == "face"
                             else process_object_mode(img, config, person_dir, count)
                             if mode == "object"
@@ -236,10 +297,12 @@ def main() -> None:
 
         jobs = interactive_configure(people)
 
-        if jobs and Confirm.ask(f"\nReady to process {sum(j['limit'] for j in jobs)} images?"):
+        if jobs:
+            _show_preview(jobs)
+            if Confirm.ask(f"Ready to process {sum(j['limit'] for j in jobs)} images?"):
                 execute_jobs(jobs)
                 rprint("\n[bold green]Done! Happy Training.[/bold green]")
-        elif not jobs:
+        else:
             rprint("[yellow]No jobs configured.[/yellow]")
 
     except KeyboardInterrupt:

if_curator/config.py

@@ -24,7 +24,21 @@ class Config:
     API_KEY: str | None = None
     OUTPUT_DIR: str = "./frigate_train"
     YEARS_FILTER: int = 10
-    MIN_FACE_WIDTH: int = 50
+
+    # Quality filtering
+    MIN_FACE_WIDTH: int = 100
+    BLUR_THRESHOLD: float = 100.0
+    MIN_CONFIDENCE: float = 0.7
+    MAX_AUTO_IMAGES: int = 80
+
+    # Output quality
+    FACE_MARGIN: float = 0.15
+    USE_FULL_RESOLUTION: bool = True
+    ENABLE_FACE_ALIGNMENT: bool = True
+
+    # Caching (opt-in to avoid unexpected files)
+    ENABLE_CACHE: bool = False
+    CACHE_DIR: str = ".if_cache"
 
     def __new__(cls) -> "Config":
         if cls._instance is None:

if_curator/diversity.py

@@ -1,9 +1,12 @@
 """
 Diversity selection for training data curation.
 
-Uses Farthest Point Sampling (FPS) algorithm with embeddings:
+Selection pipeline:
-- Faces: InsightFace embeddings
+1. Concurrent thumbnail download
-- Objects: SigLIP embeddings
+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)
+5. Cluster-aware selection (K-Medoids + FPS with hard example weighting)
 """
 
 import logging
@@ -15,6 +18,7 @@ from PIL import Image
 
 from .config import Config, get_headers
 from .embeddings import get_embedding, is_embedding_available
+from .quality import assess_quality
 
 logger = logging.getLogger(__name__)
 
@@ -28,7 +32,7 @@ def select_diverse_assets(
     progress_callback=None,
 ) -> list:
     """
-    Select diverse assets using Farthest Point Sampling or time spread.
+    Select diverse assets using cluster-aware FPS or time spread.
 
     Args:
         assets: List of asset dicts from Immich API
@@ -61,6 +65,11 @@ def select_diverse_assets(
         return _select_time_spread(assets, limit)
 
 
+# =============================================================================
+# Thumbnail & Metadata Helpers
+# =============================================================================
+
+
 def _fetch_thumbnail(asset_id: str, timeout: int = 10) -> Image.Image | None:
     """Fetch thumbnail from Immich API."""
     try:
@@ -71,13 +80,106 @@ def _fetch_thumbnail(asset_id: str, timeout: int = 10) -> Image.Image | None:
         return None
 
 
+def _get_face_bbox(asset: dict) -> tuple[float, float, float, float] | None:
+    """Extract face bounding box from asset metadata if available."""
+    for person in asset.get("people", []):
+        faces = person.get("faces", [])
+        if faces:
+            f = faces[0]
+            return (
+                f.get("boundingBoxX1", 0),
+                f.get("boundingBoxY1", 0),
+                f.get("boundingBoxX2", 0),
+                f.get("boundingBoxY2", 0),
+            )
+    return None
+
+
+def _get_face_confidence(asset: dict) -> float | None:
+    """Extract face detection confidence from asset metadata if available."""
+    for person in asset.get("people", []):
+        faces = person.get("faces", [])
+        if faces:
+            return faces[0].get("score") or faces[0].get("confidence")
+    return None
+
+
+def _crop_face_from_thumbnail(
+    img: Image.Image,
+    asset: dict,
+    margin: float = 0.25,
+) -> Image.Image | None:
+    """Crop the face region from a thumbnail using Immich bbox metadata.
+
+    By cropping before embedding, we guarantee InsightFace embeds the
+    correct person's face (not the largest face in a group photo).
+
+    Args:
+        img: Full preview thumbnail
+        asset: Asset dict with people/faces metadata
+        margin: Extra margin around the bbox (fraction, default 25%)
+
+    Returns:
+        Cropped face PIL image, or None if no face metadata available
+    """
+    bbox = _get_face_bbox(asset)
+    if bbox is None:
+        return None
+
+    x1, y1, x2, y2 = bbox
+    img_w, img_h = img.size
+
+    # Get metadata dimensions to scale bbox
+    for person in asset.get("people", []):
+        faces = person.get("faces", [])
+        if faces:
+            meta_w = faces[0].get("imageWidth") or img_w
+            meta_h = faces[0].get("imageHeight") or img_h
+            scale_x, scale_y = img_w / meta_w, img_h / meta_h
+            x1, y1 = x1 * scale_x, y1 * scale_y
+            x2, y2 = x2 * scale_x, y2 * scale_y
+            break
+
+    face_w, face_h = x2 - x1, y2 - y1
+
+    # Add margin so InsightFace's internal alignment has context
+    mx, my = face_w * margin, face_h * margin
+    crop = img.crop(
+        (
+            max(0, x1 - mx),
+            max(0, y1 - my),
+            min(img_w, x2 + mx),
+            min(img_h, y2 + my),
+        )
+    )
+
+    # Skip if too small for meaningful embedding
+    if crop.width < 30 or crop.height < 30:
+        return None
+
+    return crop
+
+
+# =============================================================================
+# Embedding Collection
+# =============================================================================
+
+
 def _select_by_embedding(
     assets: list,
     limit: int | str,
     entity_type: str,
     progress_callback=None,
 ) -> list:
-    """Select assets using embedding-based Farthest Point Sampling."""
+    """Select assets using embedding-based cluster-aware FPS.
+
+    Pipeline:
+    1. Concurrent thumbnail download
+    2. Quality filtering
+    3. Face crop extraction (face mode only)
+    4. Embedding computation
+    5. Cluster-aware selection with hard example weighting
+    """
     # Determine candidate pool (cap at 3000 for performance)
     effective_limit = 30 if limit == "auto" else limit
     pool_size = min(3000, max(effective_limit * 20, len(assets)))
@@ -89,24 +191,68 @@ def _select_by_embedding(
     else:
         candidates = assets
 
-    # Compute embeddings
+    # --- Phase 1: Concurrent thumbnail download ---
-    embeddings, valid_candidates = [], []
+    from concurrent.futures import ThreadPoolExecutor, as_completed
-    for i, asset in enumerate(candidates):
+
+    thumbnail_map: dict[str, Image.Image] = {}
+    with ThreadPoolExecutor(max_workers=8) as pool:
+        futures = {pool.submit(_fetch_thumbnail, a["id"]): a for a in candidates}
+        for i, future in enumerate(as_completed(futures)):
             if progress_callback:
                 progress_callback(i, len(candidates))
+            asset = futures[future]
+            try:
+                img = future.result()
+                if img is not None:
+                    thumbnail_map[asset["id"]] = img
+            except Exception:
+                continue
 
-        img = _fetch_thumbnail(asset["id"])
+    # --- Phase 2-4: Quality filter → Crop → Embed ---
+    embeddings, valid_candidates, confidence_scores = [], [], []
+    quality_filtered = 0
+
+    for asset in candidates:
+        img = thumbnail_map.get(asset["id"])
         if img is None:
             continue
 
-        emb = get_embedding(img, entity_type)
+        confidence = _get_face_confidence(asset)
+
+        # Quality gate: filter before expensive embedding computation
+        if entity_type == "face":
+            face_bbox = _get_face_bbox(asset)
+            quality = assess_quality(
+                img,
+                face_bbox=face_bbox,
+                confidence=confidence,
+                blur_threshold=Config.BLUR_THRESHOLD,
+                min_face_px=Config.MIN_FACE_WIDTH,
+                min_confidence=Config.MIN_CONFIDENCE,
+            )
+            if not quality.passed:
+                quality_filtered += 1
+                logger.debug(f"Quality filtered {asset['id']}: {quality.reason}")
+                continue
+
+            # Crop the target person's face before embedding
+            face_crop = _crop_face_from_thumbnail(img, asset)
+            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"])
         if emb is not None:
             embeddings.append(emb)
             valid_candidates.append(asset)
+            confidence_scores.append(confidence)
 
     if progress_callback:
         progress_callback(len(candidates), len(candidates))
 
+    if quality_filtered > 0:
+        logger.info(f"Quality filtering removed {quality_filtered} images.")
+
     if not embeddings:
         logger.warning("No valid embeddings found. Falling back to time spread.")
         return _select_time_spread(assets, limit)
@@ -115,58 +261,212 @@ def _select_by_embedding(
         logger.warning(f"Only {len(valid_candidates)} valid embeddings. Returning all.")
         return valid_candidates
 
-    # Farthest Point Sampling with vectorized distance computation
+    # --- Phase 5: Cluster-aware selection ---
-    return _farthest_point_sampling(
+    return _cluster_aware_selection(
-        embeddings, valid_candidates, limit, auto_threshold=0.15
+        embeddings,
+        valid_candidates,
+        limit,
+        entity_type=entity_type,
+        confidence_scores=confidence_scores,
     )
 
 
-def _farthest_point_sampling(
+# =============================================================================
+# K-Medoids (Lightweight Implementation)
+# =============================================================================
+
+
+def _kmedoids(dist_matrix: np.ndarray, k: int, max_iter: int = 50) -> tuple[list[int], np.ndarray]:
+    """Lightweight K-Medoids clustering using cosine distance matrix.
+
+    Args:
+        dist_matrix: (N, N) pairwise distance matrix
+        k: Number of clusters
+        max_iter: Maximum iterations for swap step
+
+    Returns:
+        (medoid_indices, cluster_labels) tuple
+    """
+    n = dist_matrix.shape[0]
+    rng = np.random.default_rng(42)
+
+    # Initialize medoids: first = most central point, rest = farthest from chosen
+    total_dist = dist_matrix.sum(axis=1)
+    medoids = [int(np.argmin(total_dist))]
+
+    for _ in range(k - 1):
+        dists_to_chosen = dist_matrix[:, medoids].min(axis=1)
+        dists_to_chosen[medoids] = -np.inf
+        medoids.append(int(np.argmax(dists_to_chosen)))
+
+    # 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))
+
+    for _ in range(max_iter):
+        improved = False
+        # Try swapping each medoid with a random non-medoid
+        non_medoids = [i for i in range(n) if i not in medoids]
+        if not non_medoids:
+            break
+
+        for m_idx in range(k):
+            candidates = rng.choice(non_medoids, size=min(10, len(non_medoids)), replace=False)
+            for cand in candidates:
+                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))
+                if new_cost < cost:
+                    medoids = new_medoids
+                    labels = new_labels
+                    cost = new_cost
+                    improved = True
+                    break
+            if improved:
+                break
+
+        if not improved:
+            break
+
+    return medoids, labels
+
+
+# =============================================================================
+# Cluster-Aware Selection (K-Medoids + FPS Hybrid)
+# =============================================================================
+
+
+def _compute_adaptive_threshold(emb_normed: np.ndarray, entity_type: str) -> 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.
+    """
+    n = len(emb_normed)
+    sample_size = min(200, n)
+    rng = np.random.default_rng(42)
+    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)]
+    median_dist = float(np.median(upper_tri))
+
+    # Faces: 20% of median (tighter — want fewer, more distinct images)
+    # 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.info(
+        f"Adaptive threshold: {threshold:.4f} "
+        f"(median_dist={median_dist:.4f}, fraction={fraction}, type={entity_type})"
+    )
+    return threshold
+
+
+def _cluster_aware_selection(
     embeddings: list,
     candidates: list,
     limit: int | str,
-    auto_threshold: float = 0.15,
+    entity_type: str = "face",
+    confidence_scores: list | None = None,
 ) -> list:
-    """Vectorized Farthest Point Sampling."""
+    """Two-stage selection: K-Medoids clustering → FPS with hard example weighting.
+
+    Stage 1: Cluster embeddings into k groups, select medoids as initial picks.
+             This guarantees at least one representative from every distinct "look".
+
+    Stage 2: Fill remaining budget with FPS across cluster boundaries,
+             biasing toward hard examples (low-confidence candidates).
+    """
     emb_matrix = np.vstack(embeddings)  # (N, D)
     n = len(emb_matrix)
 
-    # Normalize for cosine distance (cosine_dist = 1 - cosine_sim)
+    # Normalize for cosine distance
     norms = np.linalg.norm(emb_matrix, axis=1, keepdims=True)
     emb_normed = emb_matrix / np.maximum(norms, 1e-8)
 
-    # Start from median-time sample
+    # Build confidence weight array for hard example boosting
-    selected = [n // 2]
+    conf_array = np.ones(n)
+    if confidence_scores and entity_type == "face":
+        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
+    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
+    logger.info(f"Clustering {n} embeddings into {k} groups (K-Medoids)...")
+
+    # Compute full cosine distance matrix
+    dist_matrix = 1 - emb_normed @ emb_normed.T
+
+    medoid_indices, cluster_labels = _kmedoids(dist_matrix, k)
+    selected = list(medoid_indices)
+    selected_set = set(selected)
+
+    logger.info(f"Selected {len(selected)} cluster medoids as initial picks.")
+
+    # --- Stage 2: FPS with hard example weighting ---
     min_dists = np.full(n, np.inf)
 
-    target = 500 if limit == "auto" else limit
+    # Initialize min distances from all medoids
+    for idx in selected:
+        dists = dist_matrix[idx]
+        min_dists = np.minimum(min_dists, dists)
+    for idx in selected:
+        min_dists[idx] = -np.inf
 
     while len(selected) < target:
-        # Update min distances with last selected point
+        # Hard example weighting: boost distance for low-confidence candidates
-        last_emb = emb_normed[selected[-1]]
+        # Confidence < 0.85 gets up to 1.5× distance boost
-        dists_to_last = 1 - emb_normed @ last_emb  # Cosine distance
+        hard_weight = np.where(conf_array < 0.85, 1.0 + (0.85 - conf_array) * 2.0, 1.0)
-        min_dists = np.minimum(min_dists, dists_to_last)
+        weighted_dists = min_dists * hard_weight
-        min_dists[selected[-1]] = -np.inf  # Exclude already selected
 
-        # Find farthest point
+        best_idx = int(np.argmax(weighted_dists))
-        best_idx = np.argmax(min_dists)
+        best_dist = min_dists[best_idx]  # Use unweighted for threshold comparison
-        best_dist = min_dists[best_idx]
 
         if best_dist == -np.inf:
             break  # All points selected
 
         if limit == "auto" and best_dist < auto_threshold:
             logger.info(
-                f"Auto-stop: Next best image {best_dist:.3f} away (threshold {auto_threshold})."
+                f"Auto-stop: Next best image {best_dist:.3f} away " f"(adaptive threshold {auto_threshold:.4f})."
             )
             break
 
         selected.append(best_idx)
+        selected_set.add(best_idx)
+
+        # Update min distances
+        dists_to_new = dist_matrix[best_idx]
+        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(
+            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.")
 
-    logger.info(f"Smart selection complete. Picked {len(selected)} diverse images.")
     return [candidates[i] for i in selected]
 
 
+# =============================================================================
+# Time Spread Fallback
+# =============================================================================
+
+
 def _select_time_spread(assets: list, limit: int | str) -> list:
     """Select N assets evenly distributed in time."""
     if limit == "auto":

if_curator/embeddings.py

@@ -1,8 +1,9 @@
 """
 Unified embedding interface for faces and objects.
 
-- Faces: InsightFace (ArcFace/Buffalo_L)
+- Faces: InsightFace (ArcFace/Buffalo_L) — or reuse from Immich
 - Objects: SigLIP (Vision Transformer via transformers)
+- Caching: Disk-based cache avoids recomputation on reruns
 """
 
 import contextlib
@@ -14,6 +15,8 @@ import cv2
 import numpy as np
 from PIL import Image
 
+from .cache import get_cache
+
 logger = logging.getLogger(__name__)
 
 # Lazy-loaded singletons
@@ -48,8 +51,10 @@ def get_insightface_app():
 
         # Determine device: 0 for GPU, -1 for CPU
         gpu_providers = {
-            "CUDAExecutionProvider", "ROCmExecutionProvider",
+            "CUDAExecutionProvider",
-            "MPSExecutionProvider", "CoreMLExecutionProvider",
+            "ROCmExecutionProvider",
+            "MPSExecutionProvider",
+            "CoreMLExecutionProvider",
         }
         ctx_id = -1 if _is_force_cpu() else (0 if gpu_providers & set(providers) else -1)
 
@@ -73,6 +78,7 @@ def get_insightface_app():
             logger.warning("Retrying InsightFace on CPU...")
             try:
                 from insightface.app import FaceAnalysis
+
                 _insightface_app = FaceAnalysis(name="buffalo_l", root="~/.insightface")
                 _insightface_app.prepare(ctx_id=-1, det_size=(640, 640))
                 return _insightface_app
@@ -179,14 +185,83 @@ def get_object_embedding(img_pil: Image.Image) -> np.ndarray | None:
         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
+# Unified Interface with Caching
 # =============================================================================
 
 
-def get_embedding(img_pil: Image.Image, entity_type: str = "face") -> np.ndarray | None:
+def get_embedding(
-    """Get embedding for an image based on entity type ('face' or 'object')."""
+    img_pil: Image.Image,
-    return get_face_embedding(img_pil) if entity_type == "face" else get_object_embedding(img_pil)
+    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.
+
+    Priority:
+    1. Pre-fetched Immich embedding (if provided)
+    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
+    model_key = "immich" 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, model_key)
+        return immich_embedding
+
+    # 2. Check disk cache
+    if cache:
+        cached = cache.get(asset_id, model_key)
+        if cached is not None:
+            return cached
+
+    # 3. Compute locally
+    if entity_type == "face":
+        emb = get_face_embedding(img_pil)
+        model_key = "insightface"
+    else:
+        emb = get_object_embedding(img_pil)
+
+    # Cache the result
+    if emb is not None and cache:
+        cache.put(asset_id, emb, model_key)
+
+    return emb
 
 
 def is_embedding_available(entity_type: str = "face") -> bool:

if_curator/image_processing.py

@@ -3,8 +3,11 @@
 import logging
 import os
 
+import numpy as np
 from PIL import Image
 
+from .config import Config
+
 logger = logging.getLogger(__name__)
 
 # Lazy singleton
@@ -16,20 +19,59 @@ def get_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).
+
+    This produces a normalized face crop that matches exactly what ArcFace
+    was trained on, improving recognition accuracy.
+
+    Args:
+        img: Full PIL image containing the face
+        landmarks: 5-point facial landmarks [[x,y], ...] (eyes, nose, mouth corners)
+
+    Returns:
+        Aligned 112x112 face image, or None if alignment fails
+    """
+    try:
+        from insightface.utils.face_align import norm_crop
+
+        img_np = np.asarray(img)
+        lm = np.array(landmarks, dtype=np.float32)
+        if lm.shape != (5, 2):
+            logger.debug(f"Invalid landmark shape: {lm.shape}, expected (5, 2)")
+            return None
+        aligned = norm_crop(img_np, lm)
+        return Image.fromarray(aligned)
+    except ImportError:
+        logger.debug("InsightFace not available for face alignment")
+        return None
+    except Exception as e:
+        logger.debug(f"Face alignment failed: {e}")
+        return None
+
+
 def process_face_mode(
     img: Image.Image,
     asset: dict,
     person: dict,
     output_dir: str,
     count: int,
-    min_width: int = 50,
+    min_width: int | None = None,
 ) -> bool:
-    """Crop face based on Immich metadata and save to output directory."""
+    """Crop face based on Immich metadata and save to output directory.
+
+    If face alignment is enabled and landmarks are available, produces
+    an aligned 112x112 crop. Otherwise falls back to bounding box crop
+    with configurable margin.
+    """
+    min_width = min_width or Config.MIN_FACE_WIDTH
+
     # Find face metadata for this person
     face_info = None
     for p in asset.get("people", []):
@@ -57,8 +99,20 @@ def process_face_mode(
         logger.debug(f"Face too small ({face_w:.1f}x{face_h:.1f})")
         return False
 
-    # Add 10% margin
+    # Try face alignment if enabled and landmarks available
-    margin_x, margin_y = face_w * 0.10, face_h * 0.10
+    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:
+                aligned.save(os.path.join(output_dir, f"{count}.jpg"), format="JPEG")
+                return True
+
+    # Fall back to bounding box crop with configurable margin
+    margin = Config.FACE_MARGIN
+    margin_x, margin_y = face_w * margin, face_h * margin
     crop_box = (
         max(0, x1 - margin_x),
         max(0, y1 - margin_y),
@@ -87,9 +141,7 @@ def process_object_mode(
 
         found = False
         for idx, (box, cls_id, conf) in enumerate(
-            (box, int(box.cls[0]), float(box.conf[0]))
+            (box, int(box.cls[0]), float(box.conf[0])) for r in results for box in r.boxes
-            for r in results
-            for box in r.boxes
         ):
             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()

if_curator/immich_api.py

@@ -1,15 +1,30 @@
-"""Immich API client for fetching people and assets."""
+"""Immich API client for fetching people, assets, and face data."""
 
 import logging
+from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
+from io import BytesIO
 
+import numpy as np
 import requests
+from PIL import Image
 
 from .config import Config, get_headers
 
 logger = logging.getLogger(__name__)
 
 
+@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
+    image_height: int
+
+
 def get_people() -> list[dict]:
     """Fetch all people from Immich."""
     try:
@@ -68,6 +83,107 @@ def fetch_all_assets(person: dict) -> list[dict]:
     return assets
 
 
+def fetch_face_data(asset_id: str, person_id: str | None = None) -> FaceData | None:
+    """Fetch pre-computed face data (embedding, bbox, confidence) from Immich.
+
+    Queries GET /api/faces?id={asset_id} to retrieve face detection results
+    that Immich already computed using InsightFace Buffalo_L.
+
+    Args:
+        asset_id: The asset to get face data for
+        person_id: Optional person ID to match the specific face
+
+    Returns:
+        FaceData with embedding, bbox, and confidence, or None if unavailable
+    """
+    try:
+        resp = requests.get(
+            f"{Config.IMMICH_URL}/api/faces",
+            params={"id": asset_id},
+            headers=get_headers(),
+            timeout=10,
+        )
+
+        if not resp.ok:
+            logger.debug(f"Face data endpoint returned {resp.status_code} for {asset_id}")
+            return None
+
+        faces = resp.json()
+        if not faces:
+            return None
+
+        # Match the target person if specified
+        face = None
+        if person_id:
+            face = next((f for f in faces if f.get("person", {}).get("id") == person_id), None)
+        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),
+            face.get("boundingBoxX2", 0),
+            face.get("boundingBoxY2", 0),
+        )
+
+        return FaceData(
+            embedding=embedding,
+            bbox=bbox,
+            confidence=face.get("score") or face.get("confidence"),
+            image_width=face.get("imageWidth", 0),
+            image_height=face.get("imageHeight", 0),
+        )
+
+    except requests.RequestException as e:
+        logger.debug(f"Failed to fetch face data for {asset_id}: {e}")
+        return None
+    except (KeyError, TypeError, ValueError) as e:
+        logger.debug(f"Failed to parse face data for {asset_id}: {e}")
+        return None
+
+
+def fetch_full_image(asset_id: str, timeout: int = 60) -> Image.Image | None:
+    """Fetch full-resolution image from Immich, falling back to preview thumbnail.
+
+    The /original endpoint may return HEIC, RAW, or video files that PIL
+    cannot open directly. In that case, we fall back to the JPEG thumbnail.
+    """
+    # Try original first
+    try:
+        resp = requests.get(
+            f"{Config.IMMICH_URL}/api/assets/{asset_id}/original",
+            headers=get_headers(),
+            timeout=timeout,
+        )
+        if resp.ok:
+            try:
+                return Image.open(BytesIO(resp.content))
+            except Exception:
+                logger.debug(f"PIL can't open original for {asset_id}, falling back to preview")
+    except requests.RequestException:
+        logger.debug(f"Original request failed for {asset_id}, falling back to preview")
+
+    # Fall back to preview thumbnail (always JPEG)
+    try:
+        resp = requests.get(
+            f"{Config.IMMICH_URL}/api/assets/{asset_id}/thumbnail?size=preview&format=JPEG",
+            headers=get_headers(),
+            timeout=30,
+        )
+        if resp.ok:
+            return Image.open(BytesIO(resp.content))
+    except Exception as e:
+        logger.error(f"Failed to fetch image {asset_id}: {e}")
+
+    return None
+
+
 def filter_recent_assets(assets: list[dict], years: int | None = None) -> list[dict]:
     """Filter assets to keep only those from the last N years."""
     years = years or Config.YEARS_FILTER

if_curator/logging.py

@@ -36,9 +36,7 @@ def setup_logging(verbose: bool = False) -> logging.Logger:
     # File handler (always debug level)
     file_handler = logging.FileHandler("immich_export.log")
     file_handler.setLevel(logging.DEBUG)
-    file_handler.setFormatter(
+    file_handler.setFormatter(logging.Formatter("%(asctime)s - %(name)s - %(levelname)s - %(message)s"))
-        logging.Formatter("%(asctime)s - %(name)s - %(levelname)s - %(message)s")
-    )
     root.addHandler(file_handler)
 
     # Silence noisy libraries

if_curator/quality.py

@@ -0,0 +1,132 @@
+"""Image quality assessment for training data curation.
+
+Filters out images that would hurt Frigate's ArcFace model training:
+blur, grayscale/IR, over/underexposure, tiny faces, low-confidence detections.
+"""
+
+import logging
+from dataclasses import dataclass, field
+
+import cv2
+import numpy as np
+from PIL import Image
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class QualityResult:
+    """Result of quality assessment on a face/image crop."""
+
+    passed: bool
+    reasons: list[str] = field(default_factory=list)
+
+    @property
+    def reason(self) -> str:
+        return "; ".join(self.reasons) if self.reasons else "OK"
+
+
+def check_blur(img_np: np.ndarray, threshold: float = 100.0) -> tuple[bool, str]:
+    """Detect blur using Laplacian variance.
+
+    Lower variance = blurrier image. ArcFace needs clear facial features.
+    """
+    gray = cv2.cvtColor(img_np, cv2.COLOR_RGB2GRAY) if img_np.ndim == 3 else img_np
+    variance = cv2.Laplacian(gray, cv2.CV_64F).var()
+    if variance < threshold:
+        return False, f"Blurry (laplacian={variance:.1f}, threshold={threshold})"
+    return True, ""
+
+
+def check_grayscale(img_np: np.ndarray) -> tuple[bool, str]:
+    """Detect grayscale/IR images by checking channel similarity.
+
+    ArcFace is trained on color images; IR/grayscale degrades recognition.
+    """
+    if img_np.ndim != 3 or img_np.shape[2] < 3:
+        return False, "Grayscale (single channel)"
+
+    # Compare channel means — IR/grayscale has nearly identical R, G, B
+    means = img_np[:, :, :3].mean(axis=(0, 1))
+    max_diff = max(abs(means[0] - means[1]), abs(means[1] - means[2]), abs(means[0] - means[2]))
+
+    if max_diff < 5.0:
+        return False, f"Grayscale/IR (channel diff={max_diff:.1f})"
+    return True, ""
+
+
+def check_exposure(img_np: np.ndarray, lo: float = 30.0, hi: float = 225.0) -> tuple[bool, str]:
+    """Check for severe under/overexposure using mean brightness.
+
+    Extremely dark or blown-out faces lack usable features.
+    """
+    gray = cv2.cvtColor(img_np, cv2.COLOR_RGB2GRAY) if img_np.ndim == 3 else img_np
+    mean_brightness = gray.mean()
+
+    if mean_brightness < lo:
+        return False, f"Underexposed (brightness={mean_brightness:.1f}, min={lo})"
+    if mean_brightness > hi:
+        return False, f"Overexposed (brightness={mean_brightness:.1f}, max={hi})"
+    return True, ""
+
+
+def check_face_size(face_width: float, face_height: float, min_px: int = 100) -> tuple[bool, str]:
+    """Validate face crop is large enough for meaningful features."""
+    if face_width < min_px or face_height < min_px:
+        return False, f"Face too small ({face_width:.0f}x{face_height:.0f}, min={min_px}px)"
+    return True, ""
+
+
+def check_confidence(score: float | None, min_conf: float = 0.7) -> tuple[bool, str]:
+    """Check detection confidence from Immich.
+
+    Low confidence often means partial, occluded, or false-positive faces.
+    """
+    if score is None:
+        return True, ""  # No score available, pass by default
+    if score < min_conf:
+        return False, f"Low confidence ({score:.2f}, min={min_conf})"
+    return True, ""
+
+
+def assess_quality(
+    img: Image.Image,
+    face_bbox: tuple[float, float, float, float] | None = None,
+    confidence: float | None = None,
+    blur_threshold: float = 100.0,
+    min_face_px: int = 100,
+    min_confidence: float = 0.7,
+) -> QualityResult:
+    """Run all quality checks on an image.
+
+    Args:
+        img: PIL Image (RGB)
+        face_bbox: (x1, y1, x2, y2) bounding box, or None to skip face-size check
+        confidence: Detection confidence score from Immich, or None
+        blur_threshold: Laplacian variance threshold for blur detection
+        min_face_px: Minimum face dimension in pixels
+        min_confidence: Minimum acceptable detection confidence
+
+    Returns:
+        QualityResult with passed=True if all checks pass
+    """
+    img_np = np.asarray(img)
+    reasons = []
+
+    # Run all checks, collect failures
+    checks = [
+        check_blur(img_np, blur_threshold),
+        check_grayscale(img_np),
+        check_exposure(img_np),
+        check_confidence(confidence, min_confidence),
+    ]
+
+    if face_bbox is not None:
+        x1, y1, x2, y2 = face_bbox
+        checks.append(check_face_size(x2 - x1, y2 - y1, min_face_px))
+
+    for passed, reason in checks:
+        if not passed:
+            reasons.append(reason)
+
+    return QualityResult(passed=len(reasons) == 0, reasons=reasons)

pyproject.toml

@@ -3,24 +3,37 @@ name = "if-curator"
 version = "0.1.0"
 description = "Immich to Frigate training sets"
 readme = "README.md"
+license = "MIT"
 requires-python = ">=3.12"
+authors = [{ name = "ds-sebastian" }]
+keywords = ["immich", "frigate", "face-recognition", "training-data", "arcface", "insightface"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Image Recognition",
+]
 dependencies = [
     "insightface>=0.7.3",
     "numpy>=2.2.6",
     "onnxruntime>=1.23.2",
-    "opencv-python>=4.12.0.88",
     "opencv-python-headless>=4.12.0.88",
     "pillow>=12.1.0",
     "python-dotenv>=1.2.1",
     "requests>=2.32.5",
     "rich>=14.2.0",
-    "scipy>=1.17.0",
-    "sentencepiece>=0.2.1",
     "torch>=2.9.1",
     "transformers>=4.57.6",
     "ultralytics>=8.3.252",
 ]
 
+[project.scripts]
+if-curator = "if_curator.cli:main"
+
+[project.urls]
+Repository = "https://github.com/ds-sebastian/if_curator"
+
 [project.optional-dependencies]
 gpu = [
     "onnxruntime-gpu>=1.23.2",
@@ -28,6 +41,7 @@ gpu = [
 
 [dependency-groups]
 dev = [
+    "pytest>=8.0",
     "ruff>=0.9.2",
 ]
 
@@ -38,3 +52,26 @@ target-version = "py312"
 [tool.ruff.lint]
 select = ["E", "F", "I"]
 
+[tool.deptry]
+pep621_dev_dependency_groups = ["dev"]
+
+[tool.deptry.package_module_name_map]
+pillow = "PIL"
+opencv-python-headless = "cv2"
+python-dotenv = "dotenv"
+insightface = "insightface"
+numpy = "numpy"
+onnxruntime = "onnxruntime"
+requests = "requests"
+rich = "rich"
+torch = "torch"
+transformers = "transformers"
+ultralytics = "ultralytics"
+
+[tool.deptry.per_rule_ignores]
+# onnxruntime-gpu is an optional dep that replaces onnxruntime at runtime
+DEP002 = ["onnxruntime-gpu"]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"