# BiRefNet TensorRT Engine

<!-- fury-evidence-20260930:start -->
## Results and visual examples

**Historical archive.** Fury's maintained BiRefNet route is the pinned PyTorch lite-matting model. The prior TensorRT execution path was retired after repeated-call alpha-matte drift; a single Torch/TRT parity image would not establish repeat determinism. These engines are not the current maintained serving route.

Available build records: [artifacts/birefnet.cutout@v1/h100-sm90/3c7d54ccd0644c963e9526b47a5b39eb9c7a045a1066ee0a41e2d19f08018b25/runtime-metadata.json](artifacts/birefnet.cutout@v1/h100-sm90/3c7d54ccd0644c963e9526b47a5b39eb9c7a045a1066ee0a41e2d19f08018b25/runtime-metadata.json), [artifacts/birefnet.cutout@v1/rtx5090-sm120/1dd326d35db80e1c5db8cba126155fd8e3cd4416053316ff3434b3bb5dede5dc/runtime-metadata.json](artifacts/birefnet.cutout@v1/rtx5090-sm120/1dd326d35db80e1c5db8cba126155fd8e3cd4416053316ff3434b3bb5dede5dc/runtime-metadata.json), [artifacts/birefnet.cutout@v1/rtx5090-sm120/3c7d54ccd0644c963e9526b47a5b39eb9c7a045a1066ee0a41e2d19f08018b25/runtime-metadata.json](artifacts/birefnet.cutout@v1/rtx5090-sm120/3c7d54ccd0644c963e9526b47a5b39eb9c7a045a1066ee0a41e2d19f08018b25/runtime-metadata.json), [artifacts/birefnet.cutout@v1/rtx5090-sm120/60133ea26f81cb9ef3d48947685bf513896113171dc7bfdcd7127333f930605c/runtime-metadata.json](artifacts/birefnet.cutout@v1/rtx5090-sm120/60133ea26f81cb9ef3d48947685bf513896113171dc7bfdcd7127333f930605c/runtime-metadata.json), [artifacts/birefnet.cutout@v1/rtx5090-sm120/9fc8590ceb69ac518e7c98eb36d5e28a4572b1a08e0af0e71d42f76c4c6fa788/runtime-metadata.json](artifacts/birefnet.cutout@v1/rtx5090-sm120/9fc8590ceb69ac518e7c98eb36d5e28a4572b1a08e0af0e71d42f76c4c6fa788/runtime-metadata.json), [metadata.json](metadata.json).

### Source notebooks and reports

- [Maintained BiRefNet runtime and regression explanation](https://github.com/love-fury/fury-backend/blob/main/ai/docs/segmentation/birefnet/README.md)

No checkpoint-matched photo gallery was located in the existing evidence reviewed for this update. The files and runtime records below are the available evidence; no visual quality claim is inferred from their presence.

### Evidence provenance

This documentation update reuses saved results; it does not rerun inference or change weights. The starting repository revision is [`fd24be593b7c`](https://huggingface.co/carpedm20/birefnet-tensorrt/tree/fd24be593b7ca6547be1ac57bd07c471476bcf60). Captions distinguish model predictions, training diagnostics, and aggregate metrics. Qualitative examples are not a representative accuracy estimate.

[Gallery sources and SHA-256 checksums](examples/20260930/provenance.json).
<!-- fury-evidence-20260930:end -->

Pre-built TensorRT engine for [BiRefNet](ZhengPeng7/BiRefNet_dynamic) background removal.

## Specifications

| Setting | Value |
|---------|-------|
| Resolution | 1024x1024 |
| Batch Size | 1-4 |
| Precision | fp16 |
| TensorRT | 10.14.1.48 |
| Built on | NVIDIA GeForce RTX 4090 |
| Engine Size | 505.64 MB |

## Usage

The engine is automatically downloaded when using BiRefNet with TensorRT enabled.

```python
from ai.models.birefnet import BiRefNetSegmentationModel

model = BiRefNetSegmentationModel(use_tensorrt=True)
result = await model.run(image, payload=payload)
```

## Rebuild

```bash
make run-ai-local -- python -m scripts.tensorrt.birefnet --force
```

Built at: 2026-01-22 07:46:24 UTC