dinnerrush/AdcSR-x4-CoreML

AdcSR ×4 for Core ML

AdcSR (CVPR 2025, "Adversarial Diffusion Compression for Real-World Image Super-Resolution") converted to Core ML (mlprogram, FP32) for use in Instant Slide.

• Input lr: 1 × 3 × 128 × 128 Float32, RGB in [-1, 1]
• Output sr: 1 × 3 × 512 × 512 Float32, RGB in [-1, 1]
• Converted with coremltools 9.0 from net_params_200.pkl and halfDecoder.ckpt; minimum deployment target macOS 15 / iOS 18
• Run on cpuAndGPU. FP16 (including the Neural Engine) produces broken output.
• Large images are processed as overlapping tiles (16 px overlap, feathered blend), then the per-channel mean/std is matched to the input.

License

The weights are derived from Stable Diffusion 2.1 and are distributed under the CreativeML Open RAIL++-M License dated November 24, 2022 (LICENSE), including its use-based restrictions (Attachment A). You must not use this model for the restricted uses, and you must require your users to comply with them. The AdcSR source code is Apache-2.0 (LICENSE-AdcSR-code).