swiftail/UltraSharpV2-Lite-onnx-coreml

4x-UltraSharpV2 Lite — ONNX re-export for the CoreML execution provider

This is not a new model. It is 4x-UltraSharpV2 Lite by Kim2091, re-exported to ONNX so that ONNX Runtime's CoreML execution provider (Apple Silicon GPU) can run the whole graph. All credit for the model goes to Kim2091.

Why a re-export

The upstream ONNX files implement RealPLKSR's partial large-kernel convolution with an in-place slice assignment, which exports as 28 ScatterND nodes. The CoreML EP can't take those, so the graph splits into ~29 CPU/CoreML partitions and memory use explodes.

This export replaces that in-place write with an equivalent split + conv + concat (export_cat.py in this repo). The math and the weights are unchanged: tiled 1080p output matches the PyTorch (MPS) result at 65.8 dB PSNR (differences only at tile borders).

Usage

• File: 4x-UltraSharpV2_Lite_cat_fp32_op17.onnx (fp32, opset 17)
• Input input: [1, 3, H, W] float32 RGB in 0..1. Output output: [1, 3, 4H, 4W]; clamp to 0..1.
• For the CoreML EP, use fixed tile shapes (e.g. 544×544 = 512 + 2×16 padding) with static input shapes; MLProgram format, CPU+GPU compute units. Roughly 1 s per tile on an M3 Pro.

Used by Mixer, a personal D&D session app, for in-app 4× upscaling.

License

Same as the original: CC BY-NC-SA 4.0 — non-commercial use only; share derivatives under the same license with attribution to Kim2091.