DAViD foreground segmentation (ViT-B/16), dynamic int8

Quantized copy of Microsoft's DAViD soft foreground segmentation model, used by Spatial-Canvas for in-browser person masks.

Upstream: https://github.com/microsoft/DAViD (ICCV 2025). Models and runtime code are MIT licensed; see LICENSE (Copyright (c) Microsoft Corporation).
Source file: foreground-segmentation-model-vitb16_384.onnx (449,231,719 bytes) from https://facesyntheticspubwedata.z6.web.core.windows.net/iccv-2025/models/.
Modification: onnxruntime.quantization.quantize_dynamic (onnxruntime 1.24.4), MatMul/Gemm weights to uint8, per-tensor. Output 180,404,556 bytes. No retraining.
Input: 1x3x512x512, BGR, [0,1]. Output: soft foreground probability.
The SynthHuman training dataset (CDLA-2.0) is not included.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support