LocalMuseAI/coreml-cyberrealistic-v8-inpaint-6bit

CyberRealistic V8 Inpaint Compact — Core ML 6-bit for LocalMuse

This repository contains a compact Core ML conversion of CyberRealistic v8.0-Inpainting by Cyberdelia for the dedicated Inpaint workflow in the LocalMuse iOS app.

The UNet and CLIP text encoder use Apple's 6-bit k-means palettization. Both VAE directions remain FP16 to preserve image encoding, decoding and mask-edge quality. This is a storage-optimized build for nominal 6 GB devices; the separate FP16 build remains the highest-quality option for 8 GB devices.

Provenance

• Official source: CyberRealistic v8.0-Inpainting, Civitai model version 1464918
• Authenticated checkpoint mirror: vclansience/SD_models
• Pinned mirror revision: a461e57d5b2d1975670b27580cc03b6d460caf2c
• Source file: cyberrealistic_v80Inpainting.safetensors
• Source size: 2132679962 bytes
• Source SHA-256: 0ee757321f40a95029c4ff7e4ae78352ed9ee67a3ddaa8d14b6b1a48161901f2
• Architecture: Stable Diffusion 1.5 Inpaint
• UNet input/output: [2, 9, 64, 64] → [2, 4, 64, 64]
• Fixed image resolution: 512×512
• UNet and text encoder storage: 6-bit palettized
• VAE encoder and decoder storage: FP16
• 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 Hugging Face mirror is used because Civitai requires an interactive account for the original download. Its LFS SHA-256 exactly matches the official Civitai API fingerprint. Pipeline configuration is pinned separately to the official Stable Diffusion 1.5 Inpaint repository at revision 8a4288a76071f7280aedbdb3253bdb9e9d5d84bb.

The package includes a monolithic sub-1 GB UNet, the matching CLIP text encoder, VAE decoder, VAE encoder and tokenizer. PROVENANCE.json records the authenticated source, conversion dependencies and SHA-256 of every compiled runtime artifact.

This package is only compatible with a true nine-channel Inpaint runtime. It must not replace a normal four-channel SD 1.5 model.

Validation

• UNet conversion parity: 74.2 dB PSNR
• CLIP text encoder conversion parity: 82.8 dB PSNR
• VAE decoder conversion parity: 61.5 dB PSNR
• VAE encoder conversion parity: 80.7 dB PSNR
• Compiled UNet input: [2, 9, 64, 64]
• Compiled UNet output: [2, 4, 64, 64]
• All four compiled components completed finite-output prediction tests
• End-to-end 30-step DPM-Solver++ Inpaint smoke test completed successfully

Recommended settings

• Scheduler: DPM-Solver++ with Karras timestep spacing
• Steps: 30 (recommended range: 10–50)
• Guidance scale: 5.0
• Resolution: 512×512
• Batch size: 1 on iOS
• Device memory class: nominal 6 GB minimum in LocalMuse

Do not attach the generic lcm-lora-sdv1-5 adapter to this checkpoint. Reference Diffusers and Core ML end-to-end tests both produce a black masked region with that combination. A true LCM variant requires an adapter distilled specifically for the nine-channel Inpaint UNet.

License and attribution

CyberRealistic is authored by Cyberdelia. The original Civitai model permits commercial use and derivative distribution. The underlying SD 1.5 terms and use-based restrictions are retained in the included CreativeML Open RAIL-M license.

The files here are modified by conversion to compiled Core ML format and by 6-bit palettization of the UNet and text encoder. No weights were retrained or merged, and no additional restrictions are imposed. This repository does not imply endorsement by Cyberdelia, Stability AI, CompVis, Runway, Apple, Hugging Face or Civitai.