| --- |
| license: bsd-3-clause |
| tags: |
| - executorch |
| - xnnpack |
| - pte |
| - on-device |
| - image-to-image |
| - super-resolution |
| --- |
| # real_esrgan_x4v3 — ExecuTorch |
|
|
| - **Source**: xinntao/Real-ESRGAN release v0.2.5.0 (realesr-general-x4v3) |
| - **License**: BSD-3-Clause |
| - **Input**: [[1, 3, 128, 128]] — RGB 0-1, 128x128 tile |
| - **Output**: SR image [1,3,512,512] RGB, nominally 0-1 but not clamped by the model — this one overshoots noticeably (measured: 5.5% of pixels outside 0-1, range -0.28 to 1.22 over ten tiles), which is ordinary ringing for a GAN upscaler. Clamp before display. |
|
|
| ## Variants |
|
|
| All variants take and return fp32 tensors — swap the `.pte` file, keep your app code. |
|
|
| | build | file | size (MB) | parity vs fp32 eager (worst corr) | Mac median (ms)* | |
| |-----------|------|-----------|------------------------------------|------------------| |
| | fp32 | `real_esrgan_x4v3_xnnpack_fp32.pte` | 4.9 | 1.000000 | 33.9 | |
| | Core ML (fp16, iOS) | `real_esrgan_x4v3_coreml_all.pte` | 2.7 | 0.999999 | 5.1 | |
|
|
|
|
| The Core ML build is the same graph lowered to Apple's Neural Engine instead of |
| XNNPACK, which is CPU-only. Measured on an iPhone 17 Pro across seven models, it |
| runs **3.5x to 13.9x faster (median 12x)** at roughly half the file size — for |
| example Depth-Anything-V2-Small at 500.8 ms against 42.7 ms, and MODNet at 81.7 ms |
| against 5.9 ms. 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: 42.3 ms). |
| |
| ### Builds that did not earn a slot |
| |
| - **fp16 is not shipped**: it comes out at 100% of the fp32 file (4.9 MB vs 4.9 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. |
| |
| ## Verification (executorch 1.4.0, torch 2.13.0) |
| |
| Parity is measured against the fp32 eager model on real image input; `corr` is |
| the correlation over all elements of each output tensor. |
| |
| | output | shape | max_abs_diff | corr | |
| |--------|-------|--------------|------| |
| | 0 | [1, 3, 512, 512] | 1.235e-05 | 1.000000 | |
| |
| XNNPACK delegate coverage (fp32): 61.3% (106/173 ops); ops left on the portable kernels: `aten.gt.Scalar` x33, `aten.where.self` x33, `aten.upsample_nearest2d.vec` x1 |
| |
| ## Conversion |
| |
| torch.export -> to_edge_transform_and_lower(partitioner) -> .pte |
| (conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models)) |
| |
| **Notes**: int8 does not run on executorch 1.4.0, and the reason is a collision between two workarounds rather than anything about this model. PReLU segfaults on the XNNPACK delegate (upstream #17559, fix #21480 not in the stable wheel), so it has to be excluded from partitioning — and a quantized graph with a portable PReLU between delegated convolutions fails shape propagation at execute. Three conv+PReLU layers reproduce it; the same graph in fp32 is corr 1.000000. Reported at https://github.com/pytorch/executorch/pull/21480 |
| |
| <!-- funnel:v1 --> |
| |
| --- |
| |
| **More models in this format:** [ExecuTorch Model Zoo](https://huggingface.co/collections/mlboydaisuke/executorch-model-zoo-6a7ff328390b63075ffeae5e) — 31 models, each with the recipe that produced it. |
| |
| **Want a different model on-device?** [Open a request](https://github.com/john-rocky/on-device-requests) — free, open weights only; the export and its measured numbers get published publicly. |
| |
| <!-- /funnel:v1 --> |
| |
