| --- |
| license: openrail++ |
| library_name: coreai |
| pipeline_tag: image-to-image |
| tags: |
| - super-resolution |
| - diffusion |
| - core-ai |
| - apple |
| - on-device |
| - adcsr |
| - stable-diffusion |
| --- |
| |
| # AdcSR ×4 Super-Resolution — Core AI |
|
|
| On-device **×4 super-resolution** with **AdcSR** ([Adversarial Diffusion Compression](https://github.com/Guaishou74851/AdcSR), |
| CVPR 2025) converted for Apple's **Core AI** stack. AdcSR compresses the one-step diffusion model |
| [OSEDiff](https://github.com/cswry/OSEDiff) into a small **diffusion-GAN**: a pruned Stable |
| Diffusion 2.1 UNet + a half-size VAE decoder, run in **one forward pass** — no iterative denoising, |
| no prompt, no noise — so it is fast and small enough to run fully on-device, including iPhone. |
|
|
| ## What it is |
|
|
| - **fp32, ~1.7 GB.** Output matches the torch reference (cosine 1.000012). fp32 because the |
| pruned SD-2.1 UNet's attention/group-norm overflow in fp16 (NaN on smooth tiles). |
| - **Image → image, one step.** Input a low-resolution tile, get a 4× tile back. No text, no noise. |
| - **456 M parameters** (pruned SD-2.1 UNet + half VAE decoder). |
| - The graph outputs the **raw** SR; AdcSR's per-image color-match is applied host-side by |
| `SuperResolver` after tiling (baking it per-tile blows up uniform tiles). |
|
|
| ## I/O contract (per tile) |
|
|
| - **input:** `lr` `[1,3,128,128]` in `[-1,1]` (a low-resolution tile). |
| - **output:** `sr` `[1,3,512,512]` in `[-1,1]` (×4), with the reference's per-image color-match baked in. |
|
|
| ## Usage (CoreAIKit) |
|
|
| ```swift |
| import CoreAIKitVision |
| |
| let sr = try await SuperResolver(model: .adcsrX4) // downloads this repo on first use |
| let big = try await sr.upscale(cgImage) // ×4; tiles any-size input + feather-blends |
| ``` |
|
|
| `SuperResolver` splits any-size input into overlapping 128-px LR windows, runs each, and blends |
| (and caps very large inputs so the result stays a reasonable size). |
|
|
| ## License & attribution |
|
|
| - **AdcSR** (method + the pruning/training code): Apache-2.0 — Bingchen Li et al., *Adversarial |
| Diffusion Compression for Real-World Image Super-Resolution*, CVPR 2025. |
| - **Weights** are derived from **Stable Diffusion 2.1** (via OSEDiff) and therefore carry the |
| **CreativeML Open RAIL++-M** license — commercial use is permitted under its use-based |
| restrictions, the same license under which Apple distributes Stable Diffusion for Core ML. |
|
|
| This Core AI conversion inherits both. See `LICENSE` (Apache-2.0, AdcSR) and the SD-2.1 OpenRAIL++-M |
| terms. |
|
|
