---
license: creativeml-openrail-m
base_model: cyberdelia/CyberIllustrious
pipeline_tag: text-to-image
tags:
- coreml
- ios
- stable-diffusion-xl
- illustrious
- photorealistic
- localmuse
---

# CyberIllustrious V12 — Core ML for LocalMuse

A production-oriented Core ML conversion of Cyberdelia's native,
non-distilled **CyberIllustrious v12.0 FP16** checkpoint for local, on-device
1024×1024 generation in LocalMuse.

## Runtime profile

- Native Illustrious/SDXL 1024×1024; this is not Lightning, LCM, Turbo, or DMD.
- 30–45 diffusion steps; LocalMuse defaults to 35.
- DPM-Solver++ Karras, CFG 5, batch size one.
- Recommended positive conditioning prefix:
  `masterpiece, best quality, high quality, ultra-detailed`.
- Two 6-bit split UNet components, each below the iOS 1 GiB component limit.
- Exact v12 conditioning layers: CLIP-L at FP16 and OpenCLIP-bigG at 8-bit.
- Checkpoint VAE decoder and encoder for Text to Image, Transform, and
  LocalMuse Face Detail, with the Illustrious latent scale `0.18215`.
- GPU-preferred execution. This exact artifact passed a complete 30-step
  1024×1024 Core ML pipeline test on GPU; its ANE execution plan failed on the
  validation Mac and is intentionally not selected by LocalMuse.
- Requires a device in the nominal 8 GiB memory class.

Palettization changes model storage, not the normal diffusion schedule. It
does not turn v12 into a reduced-step model.

## Source and reproducibility

The official source is Cyberdelia's Civitai
[`CyberIllustrious v12.0`](https://civitai.com/models/1125067?modelVersionId=3022570),
version `3022570`, file `cyberrealistic_v120.safetensors`, SHA-256
`1b0c37e09fbfd100c78cec7c9392634bcc78eb2bb4f3004cff36755c4512b9a8`.
The same exact bytes were obtained from the pinned Hugging Face transport
mirror `grq/GemCollection` at revision
`f0ce2a74f04cf813d12e515bd0b8da5a82dfa5e9`.

The checkpoint contains non-finite values only in CLIP-L's unused final
transformer layer. The converter replaces only those values with the matching
pinned official Illustrious base layer. Prompt conditioning uses the preceding
hidden state, while pooled conditioning comes from the exact v12 second
encoder. Exact inputs, revisions, hashes, component sizes, VAE-scale evidence,
and validation results are recorded in `PROVENANCE.json`.

## License and attribution

CyberIllustrious is by Cyberdelia. The source Hugging Face model card declares
`creativeml-openrail-m`, describes CreativeML Open RAIL++-M, and links the
Stable Diffusion license. The complete linked license text is included in
`LICENSE.md`. Cyberdelia's Civitai permissions for v12 allow commercial use and
derivatives. Attribution and all use-based restrictions continue to apply to
this converted artifact.