BiRefNet_lite 512×512 (browser-ready ONNX)

A 512×512 ONNX re-export of ZhengPeng7/BiRefNet_lite that actually runs in a browser — solving the OOM wall that blocks every 1024×1024 variant from loading in onnxruntime-web. Drop it in with @huggingface/transformers to get high-quality alpha mattes entirely client-side, with no server round-trip.

Used in production by Repper for per-motif matte refinement during foreground extraction.

Quickstart (transformers.js, WebGPU)

import { AutoModel, AutoProcessor, RawImage } from '@huggingface/transformers';

const model = await AutoModel.from_pretrained('studioludens/birefnet-lite-512', {
    dtype: 'fp16',     // or 'fp32'
    device: 'webgpu',  // falls back to 'wasm' on unsupported hardware
});
const processor = await AutoProcessor.from_pretrained('studioludens/birefnet-lite-512');

const image = await RawImage.read('https://example.com/photo.jpg');
const { pixel_values } = await processor(image);

const { logits } = await model({ input_image: pixel_values });
// Apply sigmoid, upscale back to original resolution, use as alpha matte.

Why this repo exists — variant comparison

The 1024×1024 ONNX variants — including onnx-community/BiRefNet_lite-ONNX — fail in every browser backend we tested:

Root cause: BiRefNet_lite's decoder produces very large intermediate tensors at 1024×1024 (multi-scale feature maps with 1024-way concatenations). The onnxruntime-web WASM heap is hardcoded at ~2–4 GB and cannot be raised at runtime, so peak working-set exceeds available memory regardless of backend or precision.

Reducing to 512×512 shrinks intermediate tensors by 4×. At 512×512 the graph also naturally uses max 7 storage buffers per shader stage, comfortably inside WebGPU's maxStorageBuffersPerShaderStage limit (10 on older Apple Silicon adapters, 16 on Chrome ≥146), so no graph surgery is needed.

For crop-level matte refinement this is a fair trade: the crop is already small, and edge quality is indistinguishable from the 1024 reference in our tests.

Variants

config.json sets transformers.js_config.dtype = "fp16" by default. Override at load time if you want fp32.

Input / output

• Input: RGB image, resized to 512×512, ImageNet normalization (mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]), rescale factor 1/255. Layout NCHW, input tensor name input_image.
• Output: Single-channel logits at 512×512. Apply sigmoid externally to get the alpha matte in [0, 1]. Resize back to original image dimensions with bilinear interpolation.

The preprocessor_config.json uses ViTFeatureExtractor, so AutoProcessor.from_pretrained(...) works out of the box.

How it was built

Export toolchain (why it's tricky)

BiRefNet uses torchvision.ops.deform_conv2d (deformable convolution), which has no canonical ONNX symbolic. Exporting cleanly is the hard part, and every "obvious" path fails:

The fix is Kazuhito00's patch to deform_conv2d_onnx_exporter (_get_tensor_dim_size stride-based fallback), which only works against PyTorch 2.0.1's legacy tracer. Newer PyTorch versions route deform_conv2d through a different export path where the patch doesn't apply.

Why Docker

Installing PyTorch 2.0.1 locally is painful — the matching wheels are EOL, pip install torch==2.0.1 tends to pull a binary incompatible with current Python / glibc / macOS, and the surrounding torchvision / transformers pins are finicky. The reliable path is a pinned Docker image:

Python 3.10
torch==2.0.1
torchvision (compatible with 2.0.1)
transformers + deform-conv2d-onnx-exporter (Kazuhito00's patched version)

Export recipe

# Build once
docker build -t birefnet-export ./docker/

# Mount HF cache and output dir, run export
docker run --rm \
  -v "$(pwd)/docker":/work \
  -v "$HOME/.cache/huggingface":/root/.cache/huggingface \
  birefnet-export python /work/export_512_patched.py

The export script:

1. Loads ZhengPeng7/BiRefNet_lite via transformers.AutoModelForImageSegmentation.
2. Applies Kazuhito00's patched deform_conv2d_onnx_exporter before calling torch.onnx.export.
3. Exports with opset=17, fixed 512×512 input shape, constant folding enabled.
4. Writes model.onnx (fp32, ~183 MB, 17,488 nodes, max 7 bindings, 80 GatherND ops).

fp16 is produced separately via onnxruntime.transformers.float16.convert_float_to_float16 applied to the fp32 export.

Validation

Pixel-by-pixel comparison against the PyTorch forward pass on reference images. The 512 export matches PyTorch exactly (zero pixel diff when both are resized to the same output resolution).

What didn't work

• Graph surgery on the 1024 model — cascading Concat/Split ops into chains of ≤8 inputs/outputs passes the WebGPU binding-limit check, but the OOM is about intermediate tensor size, not binding count.
• onnxslim optimization — collapses cascaded ops back into the originals and inflates file size.
• Newer PyTorch exporters (2.1.x, 2.6.x dynamo) — all fail to produce correct deform_conv2d. PyTorch 2.0.1 is the working configuration.
• WebNN — Chrome-only, still behind a flag, GatherND support unconfirmed, and requires bypassing transformers.js.

Differences from upstream BiRefNet_lite

• Input resolution 512×512 instead of 1024×1024 (unblocks browser inference).
• Correct deform_conv2d export via patched exporter on PyTorch 2.0.1 — output matches PyTorch reference exactly.
• fp16 variant shipped alongside fp32.
• No graph surgery — not needed at 512×512.

For full-image matting at 1024×1024, prefer the upstream PyTorch model or server-side ONNX. This export is tuned for browser deployment.

Limitations

• 512×512 input limits edge detail on large images — use on crops or smaller inputs for best results.
• Adapters with fewer than ~10 storage buffers per shader stage fall back to WASM; the model still runs, just slower.
• No INT8 quantization yet. A quantized variant could roughly halve the fp16 size but hasn't been validated.

Citation

@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},
  year={2024}
}