feat: add vision pipeline scaffolding with ONNX backend

Introduce the app/services/vision/ module with ABC interfaces, ONNX
Runtime backend, model registry, and per-task implementations:
- OpenCLIP ViT-B/32 embedder (image + text, 512-d)
- RapidOCR engine (PP-OCRv4 via ONNX, no PaddlePaddle)
- YOLOv8n object detector (raw ONNX, no ultralytics runtime)
- YuNet + SFace face processor (Apache 2.0, opencv_zoo, 128-d)
- DBSCAN face clustering helper

Add VisionSettings to config (mulita.yml + Pydantic), bootstrap_models.py
for first-boot weight downloads, models_data Docker volume, and ROCm
backend stub for future GPU acceleration.

No Celery tasks wired yet — models load but nothing invokes them.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/Dockerfile

@@ -30,7 +30,7 @@ RUN pip install --no-cache-dir -r requirements.txt
 COPY . .
 
 # Create necessary directories
-RUN mkdir -p /data/thumbs /data/db /data/proxies /app/config
+RUN mkdir -p /data/thumbs /data/db /data/proxies /data/models /app/config
 
 # Expose port
 EXPOSE 8000

backend/app/config.py

@@ -29,6 +29,42 @@ class PerformanceSettings(BaseModel):
     db_pool_size: int = 20
     db_pool_recycle: int = 3600
 
+class EmbedderSettings(BaseModel):
+    """CLIP / SigLIP embedding model settings"""
+    name: str = "openclip_vitb32"
+    batch_size: int = 8
+
+class OCRSettings(BaseModel):
+    """PaddleOCR / rapidocr settings"""
+    enabled: bool = True
+    languages: list[str] = ["en"]
+    min_confidence: float = 0.5
+
+class DetectorSettings(BaseModel):
+    """YOLOv8n object detection settings"""
+    enabled: bool = True
+    min_confidence: float = 0.35
+    max_detections: int = 50
+
+class FacesSettings(BaseModel):
+    """YuNet + SFace face detection/recognition settings"""
+    enabled: bool = True
+    min_face_size: int = 40
+    recognition_threshold: float = 0.4
+    cluster_eps: float = 0.35
+
+class VisionSettings(BaseModel):
+    """AI vision pipeline settings. Disabled when running on SQLite
+    (pgvector is required for embedding storage)."""
+    enabled: bool = True
+    backend: str = "onnx"           # "onnx" | "rocm" (future)
+    models_dir: str = "/data/models"
+    embedder: EmbedderSettings = EmbedderSettings()
+    ocr: OCRSettings = OCRSettings()
+    detector: DetectorSettings = DetectorSettings()
+    faces: FacesSettings = FacesSettings()
+    worker_concurrency: int = 2
+
 class MulitaConfig(BaseModel):
     """Main configuration from YAML file. Source roots and the discard
     workflow are owned by the database now — only operational settings
@@ -36,6 +72,7 @@ class MulitaConfig(BaseModel):
     thumbnails: ThumbnailSettings = ThumbnailSettings()
     scanner: ScannerSettings = ScannerSettings()
     performance: PerformanceSettings = PerformanceSettings()
+    vision: VisionSettings = VisionSettings()
 
 class Settings(BaseSettings):
     """Application settings"""
@@ -127,6 +164,10 @@ class Settings(BaseSettings):
     def performance(self) -> PerformanceSettings:
         return self.config.performance
 
+    @property
+    def vision(self) -> VisionSettings:
+        return self.config.vision
+
     class Config:
         env_file = ".env"
         case_sensitive = False

backend/app/services/vision/__init__.py

@@ -0,0 +1,7 @@
+"""
+Vision pipeline services — embedding, OCR, object detection, face recognition.
+
+All inference is done through the ModelRegistry singleton, which lazy-loads
+ONNX Runtime sessions on first use and caches them for the lifetime of the
+worker process.
+"""

backend/app/services/vision/base.py

@@ -0,0 +1,89 @@
+"""
+Abstract base classes for vision backends.
+
+Each ABC defines the contract a backend must satisfy. The default
+implementation is ONNXBackend (onnx_backend.py). A ROCm backend can be
+added later by subclassing these ABCs and registering via
+settings.vision.backend.
+"""
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+import numpy as np
+
+
+@dataclass
+class DetectionBox:
+    """A single object detection result."""
+    label: str
+    confidence: float
+    bbox: list[float]  # [x1, y1, x2, y2] normalized 0-1
+
+
+@dataclass
+class OCRResult:
+    """A single OCR text region."""
+    text: str
+    confidence: float
+    bbox: list[float]  # [x1, y1, x2, y2] normalized 0-1
+    language: str = ""
+
+
+@dataclass
+class FaceDetection:
+    """A detected face with its recognition embedding."""
+    bbox: list[float]  # [x1, y1, x2, y2] normalized 0-1
+    embedding: np.ndarray  # float32 vector (128-d for SFace)
+    quality: float
+
+
+class Embedder(ABC):
+    """Generates image and text embeddings (e.g. OpenCLIP ViT-B/32)."""
+
+    @abstractmethod
+    def embed_image(self, image: np.ndarray) -> np.ndarray:
+        """Return a normalized float32 embedding vector for an RGB image."""
+        ...
+
+    @abstractmethod
+    def embed_text(self, text: str) -> np.ndarray:
+        """Return a normalized float32 embedding vector for a text query."""
+        ...
+
+    @property
+    @abstractmethod
+    def dim(self) -> int:
+        """Dimensionality of the output embedding."""
+        ...
+
+
+class OCREngine(ABC):
+    """Extracts text from images (e.g. rapidocr-onnxruntime)."""
+
+    @abstractmethod
+    def run(self, image: np.ndarray) -> list[OCRResult]:
+        """Return OCR results for an RGB image."""
+        ...
+
+
+class ObjectDetector(ABC):
+    """Detects objects in images (e.g. YOLOv8n)."""
+
+    @abstractmethod
+    def detect(self, image: np.ndarray) -> list[DetectionBox]:
+        """Return detections for an RGB image."""
+        ...
+
+
+class FaceProcessor(ABC):
+    """Detects faces and extracts recognition embeddings (e.g. YuNet + SFace)."""
+
+    @abstractmethod
+    def process(self, image: np.ndarray) -> list[FaceDetection]:
+        """Return face detections with embeddings for an RGB image."""
+        ...
+
+    @property
+    @abstractmethod
+    def embedding_dim(self) -> int:
+        """Dimensionality of face embedding vectors."""
+        ...

backend/app/services/vision/bootstrap_models.py

@@ -0,0 +1,83 @@
+"""
+Download vision model weights on first worker boot.
+
+Run as:  python -m app.services.vision.bootstrap_models
+
+Or called from the vision worker entrypoint before Celery starts.
+Downloads are idempotent — existing files with matching sizes are skipped.
+"""
+import logging
+import os
+from pathlib import Path
+from urllib.request import urlretrieve
+
+from app.config import settings
+
+logger = logging.getLogger(__name__)
+
+# (relative_path, url, expected_size_bytes_approx)
+# Sizes are approximate — used only for skip-if-exists checks, not integrity.
+MODELS = [
+    # OpenCLIP ViT-B/32 — visual and textual encoders (ONNX)
+    # These must be exported manually via export_openclip.py (see below).
+    # Placeholder entries — bootstrap will warn if missing.
+    ("embed/visual.onnx", None, None),
+    ("embed/textual.onnx", None, None),
+
+    # YOLOv8n — object detection
+    # Export: `yolo export model=yolov8n.pt format=onnx imgsz=640`
+    # Placeholder — must be exported from ultralytics offline.
+    ("detect/yolov8n.onnx", None, None),
+
+    # YuNet — face detection (Apache 2.0, opencv_zoo)
+    (
+        "face/yunet.onnx",
+        "https://github.com/opencv/opencv_zoo/raw/main/models/face_detection_yunet/face_detection_yunet_2023mar.onnx",
+        233_000,
+    ),
+
+    # SFace — face recognition (Apache 2.0, opencv_zoo)
+    (
+        "face/sface.onnx",
+        "https://github.com/opencv/opencv_zoo/raw/main/models/face_recognition_sface/face_recognition_sface_2021dec.onnx",
+        37_000_000,
+    ),
+]
+
+
+def bootstrap(models_dir: str | None = None):
+    """Ensure all model files are present. Download what we can, warn about
+    files that need manual export."""
+    base = Path(models_dir or settings.vision.models_dir)
+    base.mkdir(parents=True, exist_ok=True)
+
+    for rel_path, url, expected_size in MODELS:
+        dest = base / rel_path
+        dest.parent.mkdir(parents=True, exist_ok=True)
+
+        if dest.exists():
+            logger.debug("Model already exists: %s", dest)
+            continue
+
+        if url is None:
+            logger.warning(
+                "Model file %s not found and has no auto-download URL. "
+                "See bootstrap_models.py for export instructions.",
+                dest,
+            )
+            continue
+
+        logger.info("Downloading %s → %s", url, dest)
+        try:
+            urlretrieve(url, str(dest))
+            actual = dest.stat().st_size
+            logger.info("Downloaded %s (%d bytes)", rel_path, actual)
+        except Exception as e:
+            logger.error("Failed to download %s: %s", rel_path, e)
+            if dest.exists():
+                dest.unlink()
+
+
+if __name__ == "__main__":
+    logging.basicConfig(level=logging.INFO)
+    bootstrap()

backend/app/services/vision/clustering.py

@@ -0,0 +1,42 @@
+"""
+Face embedding clustering using DBSCAN with cosine distance.
+
+Called by the periodic `recluster_faces` Celery task (PR7).
+"""
+import logging
+
+import numpy as np
+from sklearn.cluster import DBSCAN
+
+logger = logging.getLogger(__name__)
+
+
+def cluster_faces(
+    embeddings: np.ndarray,
+    eps: float = 0.35,
+    min_samples: int = 2,
+) -> np.ndarray:
+    """Cluster face embeddings using DBSCAN with cosine metric.
+
+    Args:
+        embeddings: (N, D) float32 array of L2-normalized face embeddings.
+        eps: Maximum cosine distance between two samples to be in the
+             same neighborhood. Lower = tighter clusters.
+        min_samples: Minimum cluster size.
+
+    Returns:
+        (N,) int array of cluster labels. -1 = noise / unclustered.
+    """
+    if len(embeddings) < min_samples:
+        return np.full(len(embeddings), -1, dtype=int)
+
+    db = DBSCAN(eps=eps, min_samples=min_samples, metric="cosine")
+    labels = db.fit_predict(embeddings)
+
+    n_clusters = len(set(labels) - {-1})
+    n_noise = (labels == -1).sum()
+    logger.info(
+        "Face clustering: %d embeddings → %d clusters, %d noise",
+        len(embeddings), n_clusters, n_noise,
+    )
+    return labels

backend/app/services/vision/detect.py

@@ -0,0 +1,140 @@
+"""
+YOLOv8n object detector using raw ONNX Runtime.
+
+Expects {models_dir}/detect/yolov8n.onnx, exported from ultralytics
+via bootstrap_models.py. We do NOT ship ultralytics at runtime to
+avoid dragging in torch.
+"""
+import logging
+from pathlib import Path
+
+import numpy as np
+import onnxruntime as ort
+
+from app.config import VisionSettings
+from app.services.vision.base import ObjectDetector, DetectionBox
+
+logger = logging.getLogger(__name__)
+
+_INPUT_SIZE = 640
+
+# COCO class names (80 classes)
+COCO_LABELS = [
+    "person", "bicycle", "car", "motorcycle", "airplane", "bus", "train",
+    "truck", "boat", "traffic light", "fire hydrant", "stop sign",
+    "parking meter", "bench", "bird", "cat", "dog", "horse", "sheep",
+    "cow", "elephant", "bear", "zebra", "giraffe", "backpack", "umbrella",
+    "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard",
+    "sports ball", "kite", "baseball bat", "baseball glove", "skateboard",
+    "surfboard", "tennis racket", "bottle", "wine glass", "cup", "fork",
+    "knife", "spoon", "bowl", "banana", "apple", "sandwich", "orange",
+    "broccoli", "carrot", "hot dog", "pizza", "donut", "cake", "chair",
+    "couch", "potted plant", "bed", "dining table", "toilet", "tv",
+    "laptop", "mouse", "remote", "keyboard", "cell phone", "microwave",
+    "oven", "toaster", "sink", "refrigerator", "book", "clock", "vase",
+    "scissors", "teddy bear", "hair drier", "toothbrush",
+]
+
+
+def _preprocess(image: np.ndarray) -> tuple[np.ndarray, float, float]:
+    """Letterbox-resize + normalize to NCHW float32. Returns input tensor
+    and scale factors for mapping boxes back to original coords."""
+    from PIL import Image
+
+    img = Image.fromarray(image).convert("RGB")
+    orig_w, orig_h = img.size
+
+    scale = min(_INPUT_SIZE / orig_w, _INPUT_SIZE / orig_h)
+    new_w = int(orig_w * scale)
+    new_h = int(orig_h * scale)
+    img = img.resize((new_w, new_h), Image.BICUBIC)
+
+    # Paste onto gray canvas
+    canvas = np.full((_INPUT_SIZE, _INPUT_SIZE, 3), 114, dtype=np.uint8)
+    pad_x = (_INPUT_SIZE - new_w) // 2
+    pad_y = (_INPUT_SIZE - new_h) // 2
+    canvas[pad_y : pad_y + new_h, pad_x : pad_x + new_w] = np.array(img)
+
+    blob = canvas.astype(np.float32) / 255.0
+    blob = blob.transpose(2, 0, 1)[np.newaxis]  # NCHW
+    return blob, scale, pad_x, pad_y
+
+
+def _postprocess(
+    outputs: np.ndarray,
+    scale: float,
+    pad_x: int,
+    pad_y: int,
+    orig_w: int,
+    orig_h: int,
+    conf_threshold: float,
+    max_detections: int,
+) -> list[DetectionBox]:
+    """Parse YOLOv8 output (1, 84, N) → list of DetectionBox."""
+    # outputs shape: (1, 84, N) where 84 = 4 box coords + 80 class scores
+    preds = outputs[0]  # (84, N)
+    preds = preds.T  # (N, 84)
+
+    boxes_xywh = preds[:, :4]
+    scores = preds[:, 4:]
+
+    class_ids = np.argmax(scores, axis=1)
+    confidences = scores[np.arange(len(scores)), class_ids]
+
+    mask = confidences >= conf_threshold
+    boxes_xywh = boxes_xywh[mask]
+    class_ids = class_ids[mask]
+    confidences = confidences[mask]
+
+    if len(confidences) == 0:
+        return []
+
+    # Sort by confidence, take top N
+    order = np.argsort(-confidences)[:max_detections]
+    boxes_xywh = boxes_xywh[order]
+    class_ids = class_ids[order]
+    confidences = confidences[order]
+
+    results = []
+    for i in range(len(confidences)):
+        cx, cy, w, h = boxes_xywh[i]
+        # Remove letterbox padding and rescale to original image
+        x1 = (cx - w / 2 - pad_x) / scale
+        y1 = (cy - h / 2 - pad_y) / scale
+        x2 = (cx + w / 2 - pad_x) / scale
+        y2 = (cy + h / 2 - pad_y) / scale
+        # Normalize to 0-1
+        bbox = [
+            max(0, x1 / orig_w),
+            max(0, y1 / orig_h),
+            min(1, x2 / orig_w),
+            min(1, y2 / orig_h),
+        ]
+        label = COCO_LABELS[class_ids[i]] if class_ids[i] < len(COCO_LABELS) else f"class_{class_ids[i]}"
+        results.append(DetectionBox(label=label, confidence=float(confidences[i]), bbox=bbox))
+
+    return results
+
+
+class YOLOv8Detector(ObjectDetector):
+    def __init__(self, settings: VisionSettings):
+        model_path = Path(settings.models_dir) / "detect" / "yolov8n.onnx"
+
+        opts = ort.SessionOptions()
+        opts.inter_op_num_threads = 2
+        opts.intra_op_num_threads = 2
+
+        logger.info("Loading YOLOv8n from %s", model_path)
+        self._session = ort.InferenceSession(str(model_path), opts, providers=["CPUExecutionProvider"])
+        self._conf_threshold = settings.detector.min_confidence
+        self._max_detections = settings.detector.max_detections
+
+    def detect(self, image: np.ndarray) -> list[DetectionBox]:
+        orig_h, orig_w = image.shape[:2]
+        blob, scale, pad_x, pad_y = _preprocess(image)
+        input_name = self._session.get_inputs()[0].name
+        outputs = self._session.run(None, {input_name: blob})[0]
+        return _postprocess(
+            outputs, scale, pad_x, pad_y, orig_w, orig_h,
+            self._conf_threshold, self._max_detections,
+        )

backend/app/services/vision/embed.py

@@ -0,0 +1,81 @@
+"""
+OpenCLIP ViT-B/32 embedder using ONNX Runtime.
+
+Expects two ONNX files under {models_dir}/embed/:
+  - visual.onnx   (image encoder)
+  - textual.onnx  (text encoder)
+
+These are exported from open_clip via bootstrap_models.py.
+"""
+import logging
+from pathlib import Path
+
+import numpy as np
+import onnxruntime as ort
+
+from app.config import VisionSettings
+from app.services.vision.base import Embedder
+
+logger = logging.getLogger(__name__)
+
+# OpenCLIP ViT-B/32 preprocessing constants (ImageNet norm)
+_MEAN = np.array([0.48145466, 0.4578275, 0.40821073], dtype=np.float32)
+_STD = np.array([0.26862954, 0.26130258, 0.27577711], dtype=np.float32)
+_INPUT_SIZE = 224
+
+
+def _preprocess_image(image: np.ndarray) -> np.ndarray:
+    """Resize, center-crop, normalize an RGB uint8 image to NCHW float32."""
+    from PIL import Image
+
+    img = Image.fromarray(image).convert("RGB")
+    # Resize shortest edge to _INPUT_SIZE, then center crop
+    w, h = img.size
+    scale = _INPUT_SIZE / min(w, h)
+    img = img.resize((int(w * scale), int(h * scale)), Image.BICUBIC)
+    w, h = img.size
+    left = (w - _INPUT_SIZE) // 2
+    top = (h - _INPUT_SIZE) // 2
+    img = img.crop((left, top, left + _INPUT_SIZE, top + _INPUT_SIZE))
+
+    arr = np.array(img, dtype=np.float32) / 255.0
+    arr = (arr - _MEAN) / _STD
+    arr = arr.transpose(2, 0, 1)  # HWC → CHW
+    return arr[np.newaxis]  # NCHW
+
+
+class OpenCLIPEmbedder(Embedder):
+    def __init__(self, settings: VisionSettings):
+        model_dir = Path(settings.models_dir) / "embed"
+        visual_path = model_dir / "visual.onnx"
+        textual_path = model_dir / "textual.onnx"
+
+        opts = ort.SessionOptions()
+        opts.inter_op_num_threads = 2
+        opts.intra_op_num_threads = 2
+
+        logger.info("Loading visual encoder from %s", visual_path)
+        self._visual = ort.InferenceSession(str(visual_path), opts, providers=["CPUExecutionProvider"])
+
+        logger.info("Loading textual encoder from %s", textual_path)
+        self._textual = ort.InferenceSession(str(textual_path), opts, providers=["CPUExecutionProvider"])
+
+    def embed_image(self, image: np.ndarray) -> np.ndarray:
+        inp = _preprocess_image(image)
+        input_name = self._visual.get_inputs()[0].name
+        out = self._visual.run(None, {input_name: inp})[0][0]
+        out = out / np.linalg.norm(out)
+        return out.astype(np.float32)
+
+    def embed_text(self, text: str) -> np.ndarray:
+        import open_clip
+        tokenizer = open_clip.get_tokenizer("ViT-B-32")
+        tokens = tokenizer([text]).numpy().astype(np.int64)
+        input_name = self._textual.get_inputs()[0].name
+        out = self._textual.run(None, {input_name: tokens})[0][0]
+        out = out / np.linalg.norm(out)
+        return out.astype(np.float32)
+
+    @property
+    def dim(self) -> int:
+        return 512

backend/app/services/vision/faces.py

@@ -0,0 +1,153 @@
+"""
+Face detection (YuNet) + recognition (SFace) using ONNX Runtime.
+
+Both models are from opencv_zoo (Apache 2.0 license).
+Expects {models_dir}/face/:
+  - yunet.onnx     (~75 KB)
+  - sface.onnx     (~37 MB, 128-d embeddings)
+"""
+import logging
+from pathlib import Path
+
+import numpy as np
+import onnxruntime as ort
+
+from app.config import VisionSettings
+from app.services.vision.base import FaceProcessor, FaceDetection
+
+logger = logging.getLogger(__name__)
+
+_YUNET_INPUT_SIZE = 640
+
+
+def _align_face(image: np.ndarray, landmarks: np.ndarray) -> np.ndarray:
+    """Align and crop a 112x112 face patch using 5-point landmarks.
+    landmarks shape: (5, 2) — left_eye, right_eye, nose, left_mouth, right_mouth."""
+    from PIL import Image
+    import math
+
+    left_eye = landmarks[0]
+    right_eye = landmarks[1]
+
+    dx = right_eye[0] - left_eye[0]
+    dy = right_eye[1] - left_eye[1]
+    angle = math.degrees(math.atan2(dy, dx))
+    eye_center = ((left_eye[0] + right_eye[0]) / 2, (left_eye[1] + right_eye[1]) / 2)
+    eye_dist = math.sqrt(dx * dx + dy * dy)
+    scale = 64.0 / max(eye_dist, 1e-6)  # target: eyes at ~64px apart in 112x112
+
+    img = Image.fromarray(image)
+    img = img.rotate(-angle, center=eye_center, resample=Image.BICUBIC)
+
+    cx, cy = eye_center
+    half = 56.0 / scale
+    crop = img.crop((int(cx - half), int(cy - half * 0.8), int(cx + half), int(cy + half * 1.2)))
+    crop = crop.resize((112, 112), Image.BICUBIC)
+    return np.array(crop, dtype=np.float32)
+
+
+class YuNetSFaceProcessor(FaceProcessor):
+    def __init__(self, settings: VisionSettings):
+        face_dir = Path(settings.models_dir) / "face"
+        yunet_path = face_dir / "yunet.onnx"
+        sface_path = face_dir / "sface.onnx"
+
+        opts = ort.SessionOptions()
+        opts.inter_op_num_threads = 2
+        opts.intra_op_num_threads = 2
+
+        logger.info("Loading YuNet from %s", yunet_path)
+        self._detector = ort.InferenceSession(str(yunet_path), opts, providers=["CPUExecutionProvider"])
+
+        logger.info("Loading SFace from %s", sface_path)
+        self._recognizer = ort.InferenceSession(str(sface_path), opts, providers=["CPUExecutionProvider"])
+
+        self._min_face_size = settings.faces.min_face_size
+        self._score_threshold = settings.faces.recognition_threshold
+
+    def process(self, image: np.ndarray) -> list[FaceDetection]:
+        orig_h, orig_w = image.shape[:2]
+
+        # Scale image for YuNet (expects fixed input size)
+        scale = min(_YUNET_INPUT_SIZE / orig_w, _YUNET_INPUT_SIZE / orig_h)
+        new_w = int(orig_w * scale)
+        new_h = int(orig_h * scale)
+
+        from PIL import Image as PILImage
+        resized = np.array(
+            PILImage.fromarray(image).resize((new_w, new_h), PILImage.BICUBIC),
+            dtype=np.uint8,
+        )
+
+        # YuNet expects BGR, uint8, NHWC
+        bgr = resized[:, :, ::-1].copy()
+
+        # Run detection
+        det_input = self._detector.get_inputs()[0]
+        # YuNet uses dynamic input — reshape
+        blob = bgr.astype(np.float32)[np.newaxis]  # (1, H, W, 3)
+        # Some YuNet ONNX exports expect (1, 3, H, W)
+        if det_input.shape and len(det_input.shape) == 4 and det_input.shape[1] == 3:
+            blob = blob.transpose(0, 3, 1, 2)
+
+        detections_raw = self._detector.run(None, {det_input.name: blob})
+        dets = detections_raw[0]  # (N, 15): x,y,w,h,score, 5x landmark pairs
+
+        if dets is None or len(dets) == 0:
+            return []
+
+        results = []
+        for det in dets:
+            score = float(det[4]) if len(det) > 4 else float(det[-1])
+            if score < self._score_threshold:
+                continue
+
+            x, y, w, h = det[0], det[1], det[2], det[3]
+
+            # Filter small faces
+            face_size = max(w, h) / scale
+            if face_size < self._min_face_size:
+                continue
+
+            # Rescale to original image coords
+            x1 = x / scale
+            y1 = y / scale
+            x2 = (x + w) / scale
+            y2 = (y + h) / scale
+
+            bbox = [
+                max(0, x1 / orig_w),
+                max(0, y1 / orig_h),
+                min(1, x2 / orig_w),
+                min(1, y2 / orig_h),
+            ]
+
+            # Extract landmarks (5 points) for alignment
+            if len(det) >= 15:
+                landmarks = det[5:15].reshape(5, 2) / scale
+            else:
+                # Fallback: no landmarks, skip recognition
+                continue
+
+            # Align face for recognition
+            face_crop = _align_face(image, landmarks)
+
+            # SFace expects (1, 3, 112, 112) float32, BGR, normalized
+            face_bgr = face_crop[:, :, ::-1].copy()
+            face_blob = (face_bgr / 255.0).transpose(2, 0, 1)[np.newaxis].astype(np.float32)
+
+            rec_input = self._recognizer.get_inputs()[0].name
+            embedding = self._recognizer.run(None, {rec_input: face_blob})[0][0]
+            embedding = embedding / np.linalg.norm(embedding)
+
+            results.append(FaceDetection(
+                bbox=bbox,
+                embedding=embedding.astype(np.float32),
+                quality=score,
+            ))
+
+        return results
+
+    @property
+    def embedding_dim(self) -> int:
+        return 128

backend/app/services/vision/ocr.py

@@ -0,0 +1,45 @@
+"""
+OCR engine using rapidocr-onnxruntime (PP-OCRv4 weights).
+
+No PaddlePaddle dependency — pure ONNX Runtime. Language packs are
+downloaded automatically by rapidocr on first use.
+"""
+import logging
+
+import numpy as np
+
+from app.config import VisionSettings
+from app.services.vision.base import OCREngine, OCRResult
+
+logger = logging.getLogger(__name__)
+
+
+class RapidOCREngine(OCREngine):
+    def __init__(self, settings: VisionSettings):
+        from rapidocr_onnxruntime import RapidOCR
+
+        self._min_confidence = settings.ocr.min_confidence
+        self._engine = RapidOCR()
+        logger.info("RapidOCR engine initialized")
+
+    def run(self, image: np.ndarray) -> list[OCRResult]:
+        result, _ = self._engine(image)
+        if not result:
+            return []
+
+        out = []
+        for box, text, score in result:
+            if score < self._min_confidence:
+                continue
+            # box is [[x1,y1],[x2,y2],[x3,y3],[x4,y4]] — take bounding rect
+            xs = [p[0] for p in box]
+            ys = [p[1] for p in box]
+            h, w = image.shape[:2]
+            bbox = [
+                min(xs) / w,
+                min(ys) / h,
+                max(xs) / w,
+                max(ys) / h,
+            ]
+            out.append(OCRResult(text=text, confidence=float(score), bbox=bbox))
+        return out

backend/app/services/vision/onnx_backend.py

@@ -0,0 +1,37 @@
+"""
+ONNX Runtime backend — default CPU inference for all vision models.
+
+Each create_* method returns a concrete implementation of the
+corresponding ABC from base.py. Models are loaded from ONNX files
+under settings.vision.models_dir, downloaded on first boot by
+bootstrap_models.py.
+"""
+import logging
+
+from app.config import VisionSettings
+from app.services.vision.base import Embedder, OCREngine, ObjectDetector, FaceProcessor
+
+logger = logging.getLogger(__name__)
+
+
+class ONNXBackend:
+    """Factory for ONNX Runtime-based vision model instances."""
+
+    def __init__(self, vision_settings: VisionSettings):
+        self._settings = vision_settings
+
+    def create_embedder(self) -> Embedder:
+        from app.services.vision.embed import OpenCLIPEmbedder
+        return OpenCLIPEmbedder(self._settings)
+
+    def create_ocr(self) -> OCREngine:
+        from app.services.vision.ocr import RapidOCREngine
+        return RapidOCREngine(self._settings)
+
+    def create_detector(self) -> ObjectDetector:
+        from app.services.vision.detect import YOLOv8Detector
+        return YOLOv8Detector(self._settings)
+
+    def create_face_processor(self) -> FaceProcessor:
+        from app.services.vision.faces import YuNetSFaceProcessor
+        return YuNetSFaceProcessor(self._settings)

backend/app/services/vision/registry.py

@@ -0,0 +1,77 @@
+"""
+ModelRegistry — singleton that lazy-loads vision models per worker process.
+
+Usage from Celery tasks:
+
+    from app.services.vision.registry import registry
+    embedder = registry.get_embedder()
+    vec = embedder.embed_image(img)
+
+Models are created on first access and cached for the worker's lifetime.
+The registry reads settings.vision to decide which backend to use and
+where model weights live.
+"""
+import logging
+from functools import lru_cache
+
+from app.config import settings
+from app.services.vision.base import Embedder, OCREngine, ObjectDetector, FaceProcessor
+
+logger = logging.getLogger(__name__)
+
+
+class ModelRegistry:
+    """Central access point for all vision models."""
+
+    def __init__(self):
+        self._vision = settings.vision
+
+    @lru_cache(maxsize=1)
+    def get_embedder(self) -> Embedder:
+        logger.info("Loading embedder: %s (backend=%s)", self._vision.embedder.name, self._vision.backend)
+        return self._load_backend().create_embedder()
+
+    @lru_cache(maxsize=1)
+    def get_ocr(self) -> OCREngine:
+        logger.info("Loading OCR engine (backend=%s)", self._vision.backend)
+        return self._load_backend().create_ocr()
+
+    @lru_cache(maxsize=1)
+    def get_detector(self) -> ObjectDetector:
+        logger.info("Loading object detector (backend=%s)", self._vision.backend)
+        return self._load_backend().create_detector()
+
+    @lru_cache(maxsize=1)
+    def get_face_processor(self) -> FaceProcessor:
+        logger.info("Loading face processor (backend=%s)", self._vision.backend)
+        return self._load_backend().create_face_processor()
+
+    @lru_cache(maxsize=1)
+    def _load_backend(self):
+        """Import and instantiate the configured backend."""
+        backend_name = self._vision.backend
+        if backend_name == "onnx":
+            from app.services.vision.onnx_backend import ONNXBackend
+            return ONNXBackend(self._vision)
+        elif backend_name == "rocm":
+            from app.services.vision.rocm_backend import ROCmBackend
+            return ROCmBackend(self._vision)
+        else:
+            raise ValueError(f"Unknown vision backend: {backend_name}")
+
+    def warmup(self):
+        """Pre-load all enabled models. Called from Celery worker_process_init
+        on the vision queue to avoid cold-start latency on the first task."""
+        logger.info("Warming up vision models...")
+        self.get_embedder()
+        if self._vision.ocr.enabled:
+            self.get_ocr()
+        if self._vision.detector.enabled:
+            self.get_detector()
+        if self._vision.faces.enabled:
+            self.get_face_processor()
+        logger.info("Vision model warmup complete")
+
+
+# Module-level singleton. Import this from tasks.
+registry = ModelRegistry()

backend/app/services/vision/rocm_backend.py

@@ -0,0 +1,25 @@
+"""
+ROCm backend — GPU-accelerated inference for Radeon 760M-class hardware.
+
+Stub: raises NotImplementedError on all factory methods. To enable,
+set `vision.backend: rocm` in mulita.yml once ROCm support is implemented.
+"""
+from app.config import VisionSettings
+from app.services.vision.base import Embedder, OCREngine, ObjectDetector, FaceProcessor
+
+
+class ROCmBackend:
+    def __init__(self, vision_settings: VisionSettings):
+        self._settings = vision_settings
+
+    def create_embedder(self) -> Embedder:
+        raise NotImplementedError("ROCm backend not yet implemented — use 'onnx'")
+
+    def create_ocr(self) -> OCREngine:
+        raise NotImplementedError("ROCm backend not yet implemented — use 'onnx'")
+
+    def create_detector(self) -> ObjectDetector:
+        raise NotImplementedError("ROCm backend not yet implemented — use 'onnx'")
+
+    def create_face_processor(self) -> FaceProcessor:
+        raise NotImplementedError("ROCm backend not yet implemented — use 'onnx'")

backend/requirements.txt

@@ -34,6 +34,13 @@ pyexiftool==0.5.6
 # File watching
 watchfiles==0.21.0
 
+# Vision pipeline (ONNX Runtime CPU inference)
+onnxruntime==1.17.1
+open-clip-torch==2.24.0     # tokenizer + export helper; inference via ONNX
+rapidocr-onnxruntime==1.3.22
+scikit-learn==1.4.0          # DBSCAN for face clustering
+numpy>=1.26.0,<2.0
+
 # Utilities
 pyyaml==6.0.1
 pydantic==2.5.3

docker-compose.yml

@@ -69,6 +69,7 @@ services:
       - thumbs_data:/data/thumbs
       - proxies_data:/data/proxies
       - db_data:/data/db
+      - models_data:/data/models
     environment:
       - DATABASE_URL=postgresql+asyncpg://mulita:mulita@db:5432/mulita
       - REDIS_URL=redis://redis:6379
@@ -131,3 +132,4 @@ volumes:
   db_data:
   redis_data:
   pg_data:
+  models_data:

mulita.yml

@@ -22,3 +22,28 @@ performance:
   cache_ttl: 3600
   db_pool_size: 20
   db_pool_recycle: 3600
+
+# AI vision pipeline — embedding, OCR, object detection, face recognition.
+# Runs on the dedicated `vision` Celery queue (PR4+). Set enabled: false
+# to disable all vision processing.
+vision:
+  enabled: true
+  backend: onnx            # "onnx" (CPU) | "rocm" (future GPU)
+  models_dir: /data/models
+  embedder:
+    name: openclip_vitb32
+    batch_size: 8
+  ocr:
+    enabled: true
+    languages: [en]
+    min_confidence: 0.5
+  detector:
+    enabled: true
+    min_confidence: 0.35
+    max_detections: 50
+  faces:
+    enabled: true
+    min_face_size: 40
+    recognition_threshold: 0.4
+    cluster_eps: 0.35
+  worker_concurrency: 2