---
license: mit
library_name: litert
pipeline_tag: image-segmentation
base_model: ZhengPeng7/BiRefNet_lite
tags:
- litert
- tflite
- fp32
- background-removal
- foreground-segmentation
- birefnet
---

# BiRefNet-Lite for LiteRT

FP32 LiteRT/TFLite export of BiRefNet Lite for foreground segmentation and background removal. The model returns a soft foreground mask with the final sigmoid included.

Source checkpoint: [ZhengPeng7/BiRefNet_lite](https://huggingface.co/ZhengPeng7/BiRefNet_lite), revision [`aa62cd87eafb9cc43056d08ef3615a14628b831d`](https://huggingface.co/ZhengPeng7/BiRefNet_lite/tree/aa62cd87eafb9cc43056d08ef3615a14628b831d).

## Model file

[BiRefNet-Lite-fp32.tflite](BiRefNet-Lite-fp32.tflite) · FP32 · 184.00 MB. The file hash is in [SHA256SUMS](SHA256SUMS).

The TFLite signature is `main`. All input and output tensors are `float32`; shapes are fixed. Image/video decoding, preprocessing, and application postprocessing run on the host.

| Direction | Name | Shape | Meaning |
| --- | --- | --- | --- |
| Input | `pixel_values` | `[1, 3, 1024, 1024]` | Normalized RGB, NCHW |
| Output | `foreground_probability` | `[1, 1, 1024, 1024]` | Foreground probability in [0, 1], NCHW |

## Preprocessing and outputs

Convert the image to RGB, resize to **1024 × 1024** with PIL bilinear interpolation, divide by 255, and normalize with ImageNet mean `[0.485, 0.456, 0.406]` and standard deviation `[0.229, 0.224, 0.225]`. Transpose to NCHW. Resize the output probability map to the original image size for an alpha mask.

## Run on CPU

```bash
pip install ai-edge-litert huggingface_hub numpy pillow
```

```python
import numpy as np
from huggingface_hub import hf_hub_download
from ai_edge_litert.interpreter import Interpreter
from PIL import Image

image = Image.open("image.jpg").convert("RGB")
rgb = np.asarray(image.resize((1024, 1024), Image.Resampling.BILINEAR),
                 dtype=np.float32) / 255.0
mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
pixel_values = np.ascontiguousarray(((rgb - mean) / std).transpose(2, 0, 1)[None])

model_path = hf_hub_download("litert-community/BiRefNet-Lite", "BiRefNet-Lite-fp32.tflite")
interpreter = Interpreter(model_path=model_path, num_threads=4)
interpreter.allocate_tensors()
run = interpreter.get_signature_runner("main")
probability = run(pixel_values=pixel_values)["foreground_probability"][0, 0]
# Restore the soft foreground mask to the original image dimensions.
mask = Image.fromarray(probability).resize(image.size, Image.Resampling.BILINEAR)
alpha = Image.fromarray(np.rint(np.clip(np.asarray(mask), 0, 1) * 255).astype(np.uint8))
rgba = image.convert("RGBA")
rgba.putalpha(alpha)
rgba.save("foreground.png")
```

## Source and license

BiRefNet is by Peng Zheng and collaborators. See the [project](https://github.com/ZhengPeng7/BiRefNet) and [paper](https://arxiv.org/abs/2401.03407).

This FP32 conversion uses the original pretrained weights with a fixed-shape TFLite interface. License: [MIT](LICENSE).

## Citation

```bibtex
@article{zheng2024birefnet,
  title={Bilateral Reference for High-Resolution Dichotomous Image Segmentation},
  author={Zheng, Peng and Gao, Dehong and Fan, Deng-Ping and Liu, Li and Laaksonen, Jorma and Ouyang, Wanli and Sebe, Nicu},
  journal={CAAI Artificial Intelligence Research},
  volume={3},
  pages={9150038},
  year={2024}
}
```