halfmoon-mind/still-generative-detail-4x

GenerativeDetail4x — AdcSR ×4 for Core ML (Neural Engine, int8)

A Core ML conversion of AdcSR (Adversarial Diffusion Compression for Real-World Image Super-Resolution, CVPR 2025; code, weights), a one-step diffusion ×4 super-resolution network. It is used on device by the iOS app Still (한컷) for small low-light crops and downloaded on demand.

AdcSR synthesises photographic texture. Generated detail is a prediction, not evidence of captured detail.

Files

Path	Purpose
GenerativeDetail4x.mlpackage/	ML Program, iOS 17+, FP16 compute, int8 weights in six shards below 80 MiB
manifest.json	Size and SHA-256 of every file in the package
LICENSE-MODEL	CreativeML Open RAIL++-M (governs this derivative of Stable Diffusion 2.1-base), including the use restrictions in Attachment A
LICENSE-AdcSR	Apache License 2.0 of the AdcSR code and published weights

Interface

• Input input: float32 [1, 3, 128, 128], RGB in [-1, 1]
• Output output: float32 [1, 3, 512, 512], RGB in about [-1, 1]

As in the official inference, match each output tile's per-channel mean and standard deviation to its input tile. The app uses 128 px tiles with 32 px overlap and feathered blending. Low-light video frames need a 0.9–1.2 px Gaussian pre-blur. Without it, codec block edges become crack-like artefacts.

Measured on iPhone 14 Pro (A16, iOS 18.4), Neural Engine: about 0.36–0.40 s per tile and 0.16 s to load after the first, cached compilation (about 30 s). The GPU path keeps the weights in process memory and is not recommended.

Changes from the original weights

Required notice under paragraph 4 of the CreativeML Open RAIL++-M License. The following were changed from Guaishou74851/AdcSR weight/net_params_200.pkl (SHA-256 87fc2d3b2a007ebd568065f77b1a90fe6c10570408a38b7d72b28e6a16946111) and weight/pretrained/halfDecoder.ckpt (SHA-256 60c8763a8f7e79a9956cb6e0c5dd813deb5f54ef43b3665fce880acc718e5032):

1. Converted from PyTorch to a Core ML ML Program with a fixed 128 px tile.
2. Multiplied the residual stream of VAE decoder up_blocks[2:] by 1/16. Every reader of that stream is a GroupNorm, so the output is unchanged (FP32 maximum difference 1/255). This prevents FP16 overflow (66,194 > 65,504) in the last decoder block.
3. Quantized weights to int8 (per-channel, symmetric). On Mac Neural Engine, mean output difference against the FP16 package is 0.29–0.40/255.
4. Split the weight file into six shards without changing any tensor value.

The conversion scripts are kept in the app's repository (Tools/AdcSRStudy).

License and use restrictions

This model is a Derivative of the Model Stable Diffusion 2.1-base and is distributed under the CreativeML Open RAIL++-M License (LICENSE-MODEL, copyright © 2022 Stability AI and contributors). By using it you agree to the use restrictions in Attachment A of that license. Anyone who redistributes it, or distributes a product that includes it, must pass these restrictions on as an enforceable provision and give users a copy of the license. The AdcSR code and published files are © their authors under the Apache License 2.0 (LICENSE-AdcSR). This repository is not affiliated with or endorsed by Stability AI or the AdcSR authors.

Training data notice. AdcSR was distilled from OSEDiff (Apache-2.0) on LSDIR. LSDIR's project page describes it as provided for academic research purposes. OSEDiff was trained on LSDIR and the first 10,000 images of FFHQ, whose images carry licenses that include non-commercial terms. No warranty is given that commercial use of the weights is free of claims arising from the training data. Evaluate this for your own use.

Citation

@inproceedings{chen2025adversarial,
  title={Adversarial Diffusion Compression for Real-World Image Super-Resolution},
  author={Chen, Bin and Li, Gehui and Wu, Rongyuan and Zhang, Xindong and Chen, Jie and Zhang, Jian and Zhang, Lei},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2025}
}