feat: several photos per item, front/back views, accessories and per-photo background removal

README.md

@@ -320,7 +320,7 @@ Upgrading from a release before sign-in identities (the `user_identities` table)
 | Variable | Description | Required |
 |----------|-------------|----------|
 | `BG_REMOVAL_PROVIDER` | Background removal backend: `rembg` or `http` (default: `rembg`) | No |
-| `BG_REMOVAL_MODEL` | rembg model name (default: `u2net`) | No |
+| `BG_REMOVAL_MODEL` | rembg model name (default: `birefnet-general-lite`) | No |
 | `BG_REMOVAL_URL` | URL for HTTP bg removal provider | If http |
 | `BG_REMOVAL_API_KEY` | API key for HTTP bg removal provider | No |
 | `NEXT_PUBLIC_ENABLE_IP_LOCATION_FALLBACK` | Enable IP-based approximate location when browser geolocation is denied/unavailable. Off by default (sends the user's IP to a third party). Set to `true` to enable | No |
@@ -400,12 +400,12 @@ If browser geolocation is denied or unavailable, an optional fallback can approx
 
 ### Background Removal
 
-Works out of the box: the backend image ships rembg with the default `u2net` model already downloaded.
+Works out of the box: the backend image ships rembg with the default `birefnet-general-lite` model already downloaded. It is more accurate on clothes than `u2net`, and the cut-out is cleaned afterwards: small holes the model punched into a print are filled back in, loose bits of background are dropped, and a result that lost the garment entirely keeps the original photo.
 
 <details>
 <summary>Other rembg models, or an HTTP provider</summary>
 
-Change the rembg model with `BG_REMOVAL_MODEL` (options: `isnet-general-use`, `silueta`, `u2netp`). A non-default model is downloaded the first time background removal runs, so that first run needs internet access; the compose files keep it in the `rembg_models` volume so later runs work offline. On Kubernetes, mount a PersistentVolume at `/opt/rembg/models` on the image-worker if you change the model, or it is downloaded again on every pod start. The volume must be writable by uid 1000: the manifests' `fsGroup` covers a PersistentVolumeClaim, but a `hostPath` volume needs its directory owned by that uid. The worker links the baked `u2net` into the volume on start, so switching back needs no download.
+Change the rembg model with `BG_REMOVAL_MODEL` (options: `u2net`, `isnet-general-use`, `silueta`, `u2netp`). A non-default model is downloaded the first time background removal runs, so that first run needs internet access; the compose files keep it in the `rembg_models` volume so later runs work offline. On Kubernetes, mount a PersistentVolume at `/opt/rembg/models` on the image-worker if you change the model, or it is downloaded again on every pod start. The volume must be writable by uid 1000: the manifests' `fsGroup` covers a PersistentVolumeClaim, but a `hostPath` volume needs its directory owned by that uid. The worker links the baked `birefnet-general-lite` into the volume on start, so switching back needs no download.
 
 To use an HTTP provider (e.g. [withoutbg](https://github.com/nicholasgasior/withoutbg)) instead:
 

backend/Dockerfile

@@ -21,7 +21,7 @@ RUN pip install --no-cache-dir -r requirements.txt -r requirements-extras.txt
 # Importing rembg JIT-compiles pymatting through numba for over a minute, so the bake also fills
 # the numba cache. A generic CPU target keeps those entries valid on hosts other than the builder,
 # and the cache stays writable by every uid because numba refuses a cache dir it cannot write to.
-ARG BG_REMOVAL_MODEL=u2net
+ARG BG_REMOVAL_MODEL=birefnet-general-lite
 ENV REMBG_HOME=/opt/rembg \
     NUMBA_CACHE_DIR=/opt/numba_cache \
     NUMBA_CPU_NAME=generic

backend/app/api/analytics.py

@@ -12,6 +12,7 @@
 from app.models.user import User
 from app.utils.auth import get_current_user
 from app.utils.clothing import WardrobeComposition, count_composition
+from app.utils.garment_vocabulary import CATEGORIES, TYPE_CATEGORY
 from app.utils.signed_urls import sign_optional
 
 router = APIRouter(prefix="/analytics", tags=["Analytics"])
@@ -29,6 +30,28 @@ class TypeDistribution(BaseModel):
     percentage: float
 
 
+class CategoryDistribution(BaseModel):
+    category: str
+    count: int
+    percentage: float
+
+
+def category_distribution(types: list[TypeDistribution]) -> list[CategoryDistribution]:
+    """Clothing, footwear and accessories totals; every section is listed, even when empty."""
+    counts = dict.fromkeys(CATEGORIES, 0)
+    for entry in types:
+        counts[TYPE_CATEGORY.get(entry.type, "clothing")] += entry.count
+    total = sum(counts.values())
+    return [
+        CategoryDistribution(
+            category=category,
+            count=count,
+            percentage=round(count / total * 100, 1) if total else 0,
+        )
+        for category, count in counts.items()
+    ]
+
+
 class WearStats(BaseModel):
     id: UUID
     name: str | None
@@ -108,6 +131,7 @@ class AnalyticsResponse(BaseModel):
     wardrobe: WardrobeStats
     color_distribution: list[ColorDistribution]
     type_distribution: list[TypeDistribution]
+    category_distribution: list[CategoryDistribution] = []
     most_worn: list[WearStats]
     least_worn: list[WearStats]
     never_worn: list[WearStats]
@@ -431,6 +455,7 @@ def wear_stats_query(order_desc: bool, limit: int, never_worn: bool = False):
         wardrobe=wardrobe_stats,
         color_distribution=color_distribution,
         type_distribution=type_distribution,
+        category_distribution=category_distribution(type_distribution),
         most_worn=most_worn,
         least_worn=least_worn,
         never_worn=never_worn,