Minimal CoreAIKit example: AdcSR ×4 super-resolution on-device. Pick a photo, tap Upscale ×4.
Uses SuperResolver (CoreAIKitVision):
let sr = try await SuperResolver(model: .adcsrX4) // downloads from the Hub on first use
let big = try await sr.upscale(cgImage) // ×4, tiled + feather-blended
AdcSR (CVPR 2025) is a one-step diffusion-GAN: a pruned Stable Diffusion 2.1 UNet + half VAE decoder
run in a single forward (no prompt, no noise). The graph is one lr[1,3,128,128] → sr[1,3,512,512]
tile; SuperResolver tiles any-size input. License: Apache-2.0 code + SD-2.1 OpenRAIL++ weights
(commercial-OK). mlboydaisuke/AdcSR-CoreAI, fp16 ~870 MB (GPU output matches fp32, cos 1.000008).
xcodegen generate
open UpscaleDemo.xcodeproj
iOS app (com.coreaikit.upscaledemo). Build/run on device.