README.md

---
title: BiRefNet lite Openvino
emoji: 🏢
colorFrom: pink
colorTo: blue
sdk: gradio
sdk_version: 6.13.0
app_file: app.py
pinned: false
license: mit
short_description: BiRefNet lite Background Removal on CPU
---

Runs `ibrhr/BiRefNet-lite-openvino-xeon-w2145` with selectable OpenVINO model variants on CPU.

The Space accepts one uploaded image and returns:

- the predicted foreground mask
- a transparent PNG with the background removed
- processing time broken down by preprocessing, inference, and postprocessing
- runtime specs for the model variant, device, tensor shapes, and image size

The model picker includes all 8 OpenVINO files from the model repo:

| Variant | Resolution | Benchmark |
|---|---:|---:|
| INT8 NNCF | 1024x1024 | 1272 ms / 0.79 FPS |
| INT8 NNCF | 512x512 | 332 ms / 3.01 FPS |
| FP16 | 1024x1024 | 1419 ms / 0.70 FPS |
| FP16 | 512x512 | 366 ms / 2.73 FPS |
| FP32 | 1024x1024 | 1441 ms / 0.69 FPS |
| FP32 | 512x512 | 366 ms / 2.73 FPS |
| INT8 weight-only | 1024x1024 | 1440 ms / 0.69 FPS |
| INT8 weight-only | 512x512 | 366 ms / 2.73 FPS |

Set `MODEL_VARIANT` to change the default selection. Valid values are `int8_1024x1024`, `int8_512x512`, `fp16_1024x1024`, `fp16_512x512`, `fp32_1024x1024`, `fp32_512x512`, `int8wo_1024x1024`, and `int8wo_512x512`.