Upload README.md with huggingface_hub

README.md

@@ -10,7 +10,7 @@ tags:
 base_model:
 - schirrmacher/ormbg
 ---
-# ormbg_isnet — ExecuTorch XNNPACK
+# ormbg_isnet — ExecuTorch
 
 - **Source**: schirrmacher/ormbg
 - **License**: Apache-2.0
@@ -21,10 +21,18 @@ base_model:
 
 All variants take and return fp32 tensors — swap the `.pte` file, keep your app code.
 
-| precision | file | size (MB) | parity vs fp32 eager (worst corr) | Mac median (ms)* |
+| build | file | size (MB) | parity vs fp32 eager (worst corr) | Mac median (ms)* |
 |-----------|------|-----------|------------------------------------|------------------|
 | fp32 | `ormbg_isnet_xnnpack_fp32.pte` | 176.1 | 1.000000 | 121.7 |
 | int8 | `ormbg_isnet_xnnpack_int8.pte` | 44.3 | 0.999988 | 87.2 |
+| Core ML (fp16, iOS) | `ormbg_isnet_coreml_all.pte` | 89.0 | 0.999999 | 28.5 |
+
+
+The Core ML build is the same graph lowered to Apple's Neural Engine instead of
+XNNPACK, which is CPU-only. On an iPhone 17 Pro, Depth-Anything-V2-Small runs
+500.8 ms through XNNPACK and 42.7 ms through Core ML, at half the file size. It
+computes in fp16 and is iOS-only; the XNNPACK files stay the portable option and
+are what runs on Android.
 
 \*Mac arm64, single process, median of 10 — a reference point for relative cost
 only, not a device number (torch eager fp32 on the same machine: 375.5 ms).
@@ -33,9 +41,9 @@ only, not a device number (torch eager fp32 on the same machine: 375.5 ms).
 
 Correlation is a first filter. These are the numbers that decide:
 
-- **int8** — measured in the units that matter for this model: mask IoU at 0.5, median 0.9994 over five real images (worst 0.9916) against the fp32 build.
+- **int8** — measured in the units that matter for this model — mask IoU at 0.5: median 0.9987 over 10 real images, worst 0.9868.
 
-### Precisions that did not earn a slot
+### Builds that did not earn a slot
 
 - **fp16 is not shipped**: it comes out at 100% of the fp32 file (176.1 MB vs 176.1 MB), so it buys nothing. XNNPACK serializes convolution weights as fp32 no matter what dtype the graph carries, so on a conv-heavy model fp16 saves no disk and only adds cast operations. Reach for int8 here, not fp16.
 
@@ -52,5 +60,5 @@ XNNPACK delegate coverage (fp32): 100.0% (467/467 ops)
 
 ## Conversion
 
-torch.export -> to_edge_transform_and_lower(XnnpackPartitioner) -> .pte
+torch.export -> to_edge_transform_and_lower(partitioner) -> .pte
 (conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models))