# AdcSR ×4 Super-Resolution Lands on Apple's Core AI Stack for On-Device iPhone Upscaling

_Research · published 2026-06-25_

A CVPR 2025 image super-resolution model called AdcSR has been converted for Apple's Core AI stack, enabling 4× upscaling to run fully on-device — including on iPhone — without iterative denoising. AdcSR compresses the one-step diffusion model OSEDiff into a compact diffusion-GAN combining a pruned Stable Diffusion 2.1 UNet and a half-size VAE decoder, executing the entire upscale in a single forward pass with no text prompt or noise input. The fp32 model weighs approximately 1.7 GB and carries 456 million parameters.

Developers can access the model via CoreAIKit, where the SuperResolver class handles tiling of arbitrary-size inputs into overlapping 128-pixel low-resolution windows, runs each tile through the model, and feather-blends the results. Color matching is applied host-side after tiling to avoid artifacts on uniform tiles. The conversion inherits both the Apache-2.0 license from the AdcSR method and the CreativeML OpenRAIL++-M license from the underlying Stable Diffusion 2.1 weights, with commercial use permitted under the latter's use-based restrictions.

## Sources
- [Hugging Face Blog](https://huggingface.co/mlboydaisuke/AdcSR-CoreAI)
---
Canonical: https://genbuzz.news/posts/adcsr-x4-super-resolution-lands-on-apple-s-core-ai-stack-for-on-device-iphone-upscaling