diff --git a/recognization/README.md b/recognization/README.md
index b64e7d7..bcff2d4 100644
--- a/recognization/README.md
+++ b/recognization/README.md
@@ -25,6 +25,7 @@ Cần có đầy đủ:
 
 - `code/` - mã nguồn server, client web, client mẫu.
 - `runs/` - trọng số YOLO đã train (`runs/detect/train/weights/best.pt`).
+- `weights/dinov2_vits14_lvd142m.safetensors` - DINOv2 backbone for the identity gallery. Not in Git. See "DINOv2 weights" below.
 - `stock/` - catalog `stock.csv` và ảnh mẫu sản phẩm `sku_001` -> `sku_022`.
 - `add/` và `image.jpg` - ảnh kệ mẫu để test/demo.
 - `requirements.txt`, `requirements-dev.txt`.
@@ -49,6 +50,20 @@ source .venv/bin/activate        # Windows: .venv\Scripts\activate
 pip install -r requirements-dev.txt
 ```
 
+### DINOv2 weights
+
+The identity gallery embeds crops with DINOv2 ViT-S/14 (LVD-142M), Apache-2.0. It loads only a local file and does not contact Hugging Face, including when `YOLO_OFFLINE` is unset. Without the file, gallery identity fails and the server log says the weights are missing.
+
+The pinned file is `model.safetensors` from `timm/vit_small_patch14_dinov2.lvd142m`, sha256 `04d27f3400d059fc0cfd7d17dd1909a75bf3ea8fb3eeb48b97cb99e57ee20081` (~84 MB). Copy a cache you already have; do not fetch one from an unknown URL:
+
+```bash
+mkdir -p weights
+cp "$HOME/.cache/huggingface/hub/models--timm--vit_small_patch14_dinov2.lvd142m/snapshots/"*/model.safetensors \
+  weights/dinov2_vits14_lvd142m.safetensors
+```
+
+`DORAYAKI_DINOV2_WEIGHTS` overrides that path. A file with any other sha256 is refused.
+
 (`requirements-dev.txt` = `requirements.txt` + `httpx` cho test.)
 
 ### Bước 4: Khởi chạy recognition server