CyberRealistic Final — Core ML FP16 for LocalMuse
This repository contains an unquantized FP16 Core ML format conversion of
CyberRealistic Final by Cyberdelia
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. 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.
