LocalMuseAI/coreml-cyberrealistic-final-lcm-6bit

CyberRealistic Final LCM · Core ML 6-bit

This is a reproducible Core ML conversion prepared for the LocalMuse iOS app. It starts from CyberRealistic Final FP16, fuses the official SD 1.5 LCM-LoRA, then applies Apple's 6-bit k-means weight palettization to the UNet and CLIP text encoder. The VAE decoder and VAE encoder remain FP16.

Pinned sources

• Base: cyberdelia/CyberRealistic
  • revision: 99827f96edd717dacb28c68560680c201c55df05
  • file: CyberRealistic_FINAL_FP16.safetensors
  • bytes: 2,132,651,162
  • SHA-256: 2209c07b331a06cb28cf7c830ec758ae5b49eb97fab21f5de6b18c7be8b41554
• Adapter: latent-consistency/lcm-lora-sdv1-5
  • revision: cf2fced511dbe7e26c8d1d397e728fbab875db4b
  • file: pytorch_lora_weights.safetensors
  • bytes: 134,621,556
  • SHA-256: 8f90d840e075ff588a58e22c6586e2ae9a6f7922996ee6649a7f01072333afe4
• SD 1.5 component configuration and tokenizer: stable-diffusion-v1-5/stable-diffusion-v1-5 at 451f4fe16113bff5a5d2269ed5ad43b0592e9a14.

All source files are authenticated before the pipeline is loaded or traced. The LoRA is fused at scale 1.0 and unloaded before Core ML conversion.

Core ML layout

• Fixed resolution: 512 × 512
• Deployment target: iOS 17+
• Compute policy used for conversion: CPU and GPU
• Attention: SPLIT_EINSUM_V2
• Unet.mlmodelc: mixed FP16 / 6-bit palettized, 282 LUT operations
• TextEncoder.mlmodelc: mixed FP16 / 6-bit palettized, 74 LUT operations
• VAEDecoder.mlmodelc: FP16
• VAEEncoder.mlmodelc: FP16, included for image-to-image and Face Detail
• No safety checker is embedded
• Production payload: 957,838,366 bytes

The Core ML converter is pinned to Apple ml-stable-diffusion revision e12202c1f6405b83918b58a5d097cd61e3e1f702 with Core ML Tools 8.3.0. The conversion performed component-level PyTorch/Core ML parity checks before palettization. The final palettized package was then tested end-to-end with the LocalMuse LCM scheduler, including VAE-encoder image-to-image refinement.

Recommended settings

• Scheduler: LCM (required)
• Steps: 4–10
• Default: 8 steps
• CFG: 1.5; keep guidance in the 1.0–2.0 range
• Resolution: 512 × 512

Four, six, eight and ten steps were tested with the same prompt and seed. Eight steps gave the best quality/speed balance; ten remained stable. This tested 4–10 application range is intentionally a little wider than the adapter model card's usual 2–8 recommendation.

License and attribution

CyberRealistic and Stable Diffusion 1.5 remain subject to the CreativeML Open RAIL-M terms in LICENSE. The LCM-LoRA repository declares OpenRAIL++ and its terms are included in LCM_LORA_LICENSE.md. The use-based restrictions and attribution requirements continue to apply. This repository adds no new restrictions and does not claim authorship of the original model or adapter.

• CyberRealistic creator: Cyberdelia
• LCM-LoRA authors: Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu et al.
• Format conversion: LocalMuseAI