huochai67/wd-convnext-tagger-v3-rk3566-int8

WD ConvNeXt Tagger v3 for RK3566 INT8

INT8 RKNN conversion of SmilingWolf/wd-convnext-tagger-v3 for Rockchip RK3566 NPU inference.

Contents

• wd-convnext-tagger-v3-rk3566-int8.rknn: static-shape RK3566 model
• selected_tags.csv: matching WD v3 tag metadata

Conversion

• RKNN Toolkit2: 2.3.2
• Target: RK3566
• Input: NCHW [1, 3, 448, 448], BGR, range 0..255
• Quantization: INT8
• Calibration: 300 random images sampled from Amber-River/Pixiv-2.6M

Calibration images and the original ONNX model are not included.

RK3566 Benchmark

Measured on RK3566 with 4GB RAM, RKNN Runtime 2.3.2, and driver 0.9.8:

Metric	Result
Average NPU inference	0.709 s/image
Throughput	1.41 images/s
Peak process RSS	345 MiB

Usage

The complete conversion and board-side inference code is available at:

https://github.com/huochai67/wd-convnext-rknn-rk3566

Use selected_tags.csv from this repository with this .rknn model. Do not mix it with labels from another WD model.

License

The base model is licensed under Apache-2.0. Refer to the base model and the conversion repository for attribution and deployment details.