| --- |
| license: creativeml-openrail-m |
| tags: |
| - text-to-image |
| - stable-diffusion |
| - stable-diffusion-1.5 |
| - cyberrealistic |
| - photorealistic |
| - coreml |
| - fp16 |
| - ios |
| base_model: |
| - cyberdelia/CyberRealistic |
| pipeline_tag: text-to-image |
| library_name: ml-stable-diffusion |
| --- |
| |
| # CyberRealistic Final — Core ML FP16 for LocalMuse |
|
|
| This repository contains an unquantized FP16 Core ML format conversion of |
| [CyberRealistic Final by Cyberdelia](https://huggingface.co/cyberdelia/CyberRealistic) |
| for on-device inference in the LocalMuse iOS app. |
|
|
| ## Provenance |
|
|
| - Source repository: `cyberdelia/CyberRealistic` |
| - Pinned source revision: `99827f96edd717dacb28c68560680c201c55df05` |
| - Source file: `CyberRealistic_FINAL_FP16.safetensors` |
| - Source file size: `2,132,651,162` bytes |
| - Source SHA-256: `2209c07b331a06cb28cf7c830ec758ae5b49eb97fab21f5de6b18c7be8b41554` |
| - Architecture: Stable Diffusion 1.5 |
| - Fixed resolution: 512×512 |
| - Storage precision: FP16 |
| - Quantization/palettization: none |
| - UNet attention graph: `SPLIT_EINSUM_V2` |
| - Core ML deployment target: iOS 17 |
| - Core ML execution policy in LocalMuse: CPU + GPU |
| - Conversion tooling: Apple `ml-stable-diffusion` commit |
| `e12202c1f6405b83918b58a5d097cd61e3e1f702`, Core ML Tools 8.3.0 |
|
|
| The source checkpoint is inference-pruned FP16, not integer-quantized. The |
| UNet is split into two compiled graphs so each weight file remains below 1 GB. |
| Splitting changes only graph packaging and does not quantize the weights. The |
| CLIP text encoder, VAE decoder and VAE encoder are included, so text-to-image |
| and image-to-image/face-detail workflows are supported. |
|
|
| The model configuration and tokenizer are pinned to |
| `stable-diffusion-v1-5/stable-diffusion-v1-5` revision |
| `451f4fe16113bff5a5d2269ed5ad43b0592e9a14`. Source and configuration files |
| were authenticated by exact size and SHA-256 before conversion. |
|
|
| ## Validation |
|
|
| The conversion was checked component-by-component against the pinned PyTorch |
| source before publication: |
|
|
| - UNet Core ML parity: 73.8 dB PSNR |
| - CLIP text encoder parity: 82.7 dB PSNR |
| - VAE decoder parity: 61.7 dB PSNR |
| - VAE encoder parity: 81.7 dB PSNR |
|
|
| An end-to-end 28-step DPM-Solver++ generation completed with Apple's Swift |
| Stable Diffusion pipeline using CPU+GPU compute units. A separate image-to-image |
| generation also completed through the VAE encoder used by face detail. |
|
|
| ## Recommended settings |
|
|
| - Scheduler: DPM-Solver++ |
| - Steps: 28 |
| - Guidance scale: 7.5 |
| - Resolution: 512×512 |
| - Batch size: 1 on iOS |
| - Device memory: 8 GB minimum in LocalMuse |
|
|
| ## License and attribution |
|
|
| CyberRealistic is authored by Cyberdelia and published under the |
| [CreativeML Open RAIL-M license](LICENSE). These files are modified from the |
| original by converting the model to compiled Core ML FP16 format and splitting |
| the UNet graph. No weights were retrained and no additional restrictions are |
| imposed. The original license and its use-based restrictions continue to apply. |
|
|
| This repository does not imply endorsement by Cyberdelia, Stability AI, |
| CompVis, Runway, Apple or Hugging Face. |
|
|
