LocalMuseAI/coreml-juggernaut-xi-lightning-6bit

Juggernaut XI Lightning — LocalMuse Core ML

This repository contains the iOS-optimized Core ML conversion used by LocalMuse. It was converted directly from the exact Civitai checkpoint juggernautXL_juggXILightningByRD.safetensors for Juggernaut XL version Jugg_XI_Lightning_by_RD (model version 920957). It is not a merge and does not reuse UNet, text encoder, or VAE weights from another LocalMuse model.

Recommended generation profile

• Sampler: DPM++ SDE
• Schedule: Karras
• Steps: 6 (author range: 5–7)
• CFG: 2.0 (author range: 1.5–2.0)
• Clip skip: 2 (the SDXL converter exports the penultimate hidden layer)
• Native LocalMuse canvas: 1024×1024
• Suggested negative prompt: CGI, Unreal, Airbrushed, Digital

Optimization

• UNet: 6-bit palettization, split into two Core ML programs
• Text encoder 1: FP16
• Text encoder 2: 8-bit palettization
• VAE encoder and decoder: model-specific FP16 conversion
• Attention: SPLIT_EINSUM
• Deployment target: iOS 17
• Runtime: Neural Engine-preferred UNet and GPU VAE decode in LocalMuse

The compiled payload is fixed at 1024×1024. All large compiled weight files remain below the iOS 1 GiB per-file limit.

See PROVENANCE.md and ARTIFACT_MANIFEST.json for authenticated source and artifact hashes.

Attribution and license

Juggernaut XL is created by KandooAI / RunDiffusion. The source version is published on Civitai under the CreativeML Open RAIL++-M terms and its displayed addendum. Review the source page and LICENSE.md before redistribution or use.

• Source version: https://civitai.com/models/133005?modelVersionId=920957
• Creator/config reference: https://huggingface.co/RunDiffusion/Juggernaut-XL-Lightning